Waveform classification model training method and physiological signal waveform classification method

Abstract

The present invention relates to the technical field of physiological signal waveform classification, and in particular to a waveform classification model training method and a physiological signal waveform classification method. In the present invention, each convolution basic block of a convolution module in a neural network model comprises a grouped convolution layer and a shuffling operation layer; a convolution kernel of the grouped convolution layer uses a dilated convolution; both the dilated convolution and a grouped convolution are computationally lightweight; in addition, the convolution basic block is also provided with the shuffling operation layer; the shuffling operation layer can perform cross combination on features output by channels in an input layer of a grouped convolution, so that a final feature map output by the grouped convolution covers feature information in all the channels in the input layer, to ensure the accuracy of the features output by the convolution basic block where the grouped convolution is located, thereby ensuring the accuracy of psychological signal waveform classification.

Description

A training method for a waveform classification model and a physiological signal waveform classification method Technical Field

[0001] The present invention relates to the technical field of physiological signal waveform classification, and in particular to a waveform classification model training method and a physiological signal waveform classification method. Background Art

[0002] Physiological signal waveforms include arterial blood pressure waveforms, electrocardiogram (ECG) waveforms, and pulmonary artery pressure waveforms. By analyzing these waveforms, the subject's blood pressure can be classified. Existing technologies use machine learning models to classify blood pressure. However, these models increase blood pressure classification accuracy at the expense of increasing model parameters. Increasing model parameters increases the computational complexity of the model, thereby reducing blood pressure classification efficiency.

[0003] In summary, it is difficult for existing technologies to balance blood pressure classification efficiency and classification accuracy.

[0004] Therefore, the existing technology needs to be improved and enhanced.

[0005] Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a waveform classification model training method and a physiological signal waveform classification method, which solves the problem that the existing technology is difficult to balance blood pressure classification efficiency and classification accuracy.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] In a first aspect, the present invention provides a training method for a waveform classification model, wherein the neural network model is applied to physiological signal waveform classification, comprising:

[0009] Constructing a neural network model including a convolution module, wherein the convolution module includes a plurality of convolution basic blocks cascaded in sequence, each of the convolution basic blocks including a first grouped convolution layer, a shuffle operation layer, a second grouped convolution layer, a fully connected layer, and an output layer cascaded in sequence, wherein the convolution kernel of the first grouped convolution layer is a dilated convolution;

[0010] According to the arterial diastolic pressure and the arterial systolic pressure, a training data set is screened from a database, wherein the database is used to store training waveforms of physiological signals related to the blood pressure signal;

[0011] The neural network model is trained using the training data set to obtain a waveform classification model.

[0012] In one implementation, the training data set is screened from a database based on the arterial diastolic pressure and the arterial systolic pressure, and the database is used to store training waveforms of physiological signals related to blood pressure signals, including:

[0013] Calculating the difference between the arterial systolic pressure and the arterial diastolic pressure, which is recorded as the systolic-diastolic pressure difference;

[0014] Multiplying the systolic and diastolic pressure difference by a coefficient less than one and then adding the resultant to the arterial diastolic pressure to obtain the plateau arterial blood pressure;

[0015] According to the plateau arterial blood pressure, a training data set is screened out from the database, and the database is used to store training waveforms of physiological signals related to blood pressure signals.

[0016] In one implementation, the training data set is screened from the database based on the plateau arterial blood pressure, and the database is used to store training waveforms of physiological signals related to blood pressure signals, including:

[0017] Counting the duration of time that the plateau arterial blood pressure is continuously lower than the set blood pressure value;

[0018] Filter out the trainees whose duration is longer than the set duration from the trainees in the database and record them as target trainees;

[0019] Filtering out the training arterial pressure original waveform and / or the training electrocardiogram signal original waveform and / or the training pulmonary artery pressure original waveform related to the blood pressure signal of the target trainee and the classification training label corresponding to the target trainee from the database;

[0020] The training arterial pressure original waveform and / or the training electrocardiogram signal original waveform and / or the training pulmonary artery pressure original waveform and the classification training label are used as the training data set.

[0021] In a second aspect, an embodiment of the present invention further provides a physiological signal waveform classification method, which uses a trained neural network model determined by the above-mentioned waveform classification model training method, and the physiological signal waveform classification method includes:

[0022] Extracting original feature maps of physiological signal waveforms related to blood pressure signals;

[0023] Applying a convolution module included in a waveform classification model to the original feature map to obtain a final feature map, wherein the convolution module includes a plurality of convolution basic blocks cascaded in sequence, each of the convolution basic blocks includes a first grouped convolution layer, a shuffle operation layer, a second grouped convolution layer, a fully connected layer, and an output layer cascaded in sequence, wherein the convolution kernel of the first grouped convolution layer is a dilated convolution;

[0024] The physiological signal waveform is classified according to the final feature map to obtain a classification result.

[0025] In one implementation, the convolution basic block further includes an average pooling layer, and the average pooling layer is cascaded with the fully connected layer.

[0026] In one implementation, extracting an original feature map of a physiological signal waveform related to a blood pressure signal includes:

[0027] Determining an arterial pressure original waveform and / or an electrocardiogram signal original waveform and / or a pulmonary artery pressure original waveform related to the blood pressure signal;

[0028] Using the original waveform of arterial pressure and / or the original waveform of electrocardiogram signal and / or the original waveform of pulmonary artery pressure as the physiological signal waveform;

[0029] A feature extraction module is applied to the physiological signal waveform to obtain an original feature map, wherein the feature extraction module includes a 1x1 convolution layer, a normalization layer, and an activation function layer cascaded in sequence.

[0030] In one implementation, classifying the physiological signal waveform according to the final feature map to obtain a classification result includes:

[0031] A classification module is applied to the final feature map to classify the physiological signal waveform to obtain a classification result, wherein the classification module includes a cascaded average pooling layer and a classifier.

[0032] In a third aspect, an embodiment of the present invention further provides a training device for a waveform classification model, wherein the device includes the following components:

[0033] A model construction module is used to construct a neural network model including a convolution module, wherein the convolution module includes a plurality of convolution basic blocks cascaded in sequence, each of which includes a first group convolution layer, a shuffle operation layer, a second group convolution layer, a fully connected layer, and an output layer cascaded in sequence, wherein the convolution kernel of the first group convolution layer is a dilated convolution;

[0034] a data screening module for screening a training data set from a database based on arterial diastolic pressure and arterial systolic pressure, wherein the database is used to store training waveforms of physiological signals related to blood pressure signals;

[0035] A training module is used to train the neural network model using the training data set to obtain a waveform classification model.

[0036] In a fourth aspect, an embodiment of the present invention further provides a terminal device, wherein the terminal device includes a memory, a processor, and a physiological signal waveform classification program stored in the memory and executable on the processor, wherein when the processor executes the physiological signal waveform classification program, the steps of the physiological signal waveform classification method described above are implemented;

[0037] Alternatively, the terminal device includes a memory, a processor, and a training program for a waveform classification model stored in the memory and runnable on the processor. When the processor executes the training program for the waveform classification model, the steps of the above-mentioned training method for the waveform classification model are implemented.

[0038] In a fifth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a physiological signal waveform classification program, and when the physiological signal waveform classification program is executed by a processor, the steps of the physiological signal waveform classification method described above are implemented;

[0039] Alternatively, the computer-readable storage medium stores a training program for the waveform classification model, and when the training program for the waveform classification model is executed by the processor, the steps of the above-mentioned training method for the waveform classification model are implemented.

[0040] Beneficial effect: Each convolution basic block of the convolution module of the present invention includes a grouped convolution layer and a shuffle operation layer, wherein the convolution kernel of the grouped convolution layer adopts a dilated convolution, and both the dilated convolution and the grouped convolution are computationally lightweight, that is, the dilated convolution and the grouped convolution do not require too much computing resources to complete the optimization of features. At the same time, a shuffle operation layer is set in the convolution basic block, and the shuffle operation layer can cross-combine the features output by each channel in the input layer of the grouped convolution, so that the final feature map output by the grouped convolution covers the feature information in all channels in the input layer, thereby ensuring the accuracy of the features output by the convolution basic block where the grouped convolution is located, and further ensuring the accuracy of the psychological signal waveform classification. In summary, the present invention can both reduce the computing resources required for classification and ensure classification accuracy, that is, the present invention takes into account both the efficiency and accuracy of blood pressure classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] FIG1 is an overall flow chart of the present invention;

[0042] FIG2 is a structural diagram of a classification model in an embodiment of the present invention;

[0043] FIG3 is a structural diagram of a basic block in an embodiment of the present invention;

[0044] FIG4 is a basic block diagram of an embodiment of the present invention;

[0045] FIG5 is a schematic diagram of group convolution and shuffling operations in an embodiment of the present invention;

[0046] FIG6 is a schematic diagram of a dilated convolution in an embodiment of the present invention;

[0047] FIG7 is a schematic diagram of AHE discrimination in an embodiment of the present invention;

[0048] FIG8 is a schematic diagram of a blood pressure prediction window in an embodiment of the present invention;

[0049] FIG9 is a schematic diagram of using three waveforms to predict whether an AHE event occurs in an embodiment of the present invention;

[0050] FIG10 is a structural diagram of a physiological signal waveform classification device provided by the present invention;

[0051] FIG11 is a block diagram of the internal structure of a terminal device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following is a clear and complete description of the technical solutions of the present invention in conjunction with the embodiments and the accompanying drawings. 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.

[0053] Research has found that physiological signal waveforms include arterial blood pressure waveforms, electrocardiogram waveforms, and pulmonary artery pressure waveforms. By analyzing these waveforms, the subject's blood pressure can be classified. Existing technologies use machine learning models to classify blood pressure. However, these models improve blood pressure classification accuracy at the expense of increasing model parameters. Increasing model parameters increases the computational complexity of the model, thereby reducing blood pressure classification efficiency.

[0054] In order to solve the above technical problems, the present invention provides a training method for a waveform classification model and a physiological signal waveform classification method, which solves the problem that the existing technology is difficult to take into account both blood pressure classification efficiency and classification accuracy. In specific implementation, a neural network model is first constructed. The neural network model includes a convolution module that is cascaded in sequence. The convolution module includes a number of convolution basic blocks that are cascaded in sequence. Each of the convolution basic blocks includes a first group convolution layer, a shuffling operation layer, a second group convolution layer, a fully connected layer and an output layer that are cascaded in sequence, wherein the convolution kernel of the first group convolution layer is a hole convolution; finally, the physiological signal waveform is classified according to the final feature map to obtain a classification result. The present invention can take into account both blood pressure classification efficiency and classification accuracy.

[0055] The training method of the waveform classification model of this embodiment can be applied to a terminal device, which can be a terminal product with a waveform acquisition function, such as a blood pressure monitor. In this embodiment, as shown in FIG1 , the training method of the waveform classification model specifically includes the following steps:

[0056] S100, construct a neural network model including a convolution module, wherein the convolution module includes several convolution basic blocks cascaded in sequence, each of the convolution basic blocks includes a first group convolution layer, a shuffle operation layer, a second group convolution layer and a fully connected layer and an output layer cascaded in sequence, wherein the convolution kernel of the first group convolution layer is a dilated convolution.

[0057] S200 , screening out a training data set from a database based on the arterial diastolic pressure and the arterial systolic pressure, wherein the database is used to store training waveforms of physiological signals related to blood pressure signals.

[0058] S300: Use the training data set to train the neural network model to obtain a waveform classification model.

[0059] The structure of the neural network model is shown in Figure 2. In Figure 2, the convolution layer Convld, the normalization layer BatchNom, and the activation function layer ReLU cascaded in sequence constitute the feature extraction module, several convolution basic blocks BasicBlock cascaded in sequence constitute the convolution module, and the average pooling layer AVGPool and the classifier FCClassifier constitute the classification module.

[0060] As shown in Figure 3, each convolution basic block BasicBlock includes the first group convolution layer GConv, the shuffle operation layer Channel shuffle, the second group convolution layer GConv and the fully connected layer Concat with the output layer Dropout, and also includes the average pooling layer AVGPool.

[0061] Taking the convolution basic block BasicBlock at the head end as an example, the so-called head end is the convolution basic block BasicBlock cascaded with the activation function layer ReLU in Figure 2. The working principle of the convolution basic block BasicBlock is explained using Figure 4 as an example. The feature extraction module in Figure 2 extracts the original blood pressure feature map covered by the three waveforms: the original waveform of arterial pressure, the original waveform of electrocardiogram signal, and the original waveform of pulmonary artery pressure. These original feature maps are input to the input layer Input of the first grouped convolution layer GConv. The input layer Input of GConv groups the various features contained in the original blood pressure feature map. Figure 5 shows that two groups are divided, each with four channels. The dilated convolution shown in Figure 6 is then applied to the features in each channel. The position of the dilated convolution in Figure 5 is the connection between the input layer and the hidden layer of Figure 5 to obtain the features of each channel in the hidden layer shown in Figure 5. The scale of each dilated convolution is different. As shown in Figure 5, a shuffle operation is used to swap the channels in the hidden layer. This swapping involves swapping the channels in the hidden layer corresponding to different groups. The features in each of the swapped channels are then subjected to a normal convolution to obtain the features output by the first grouped convolution layer GConv. The second grouped convolution layer GConv is then applied to the features output by the first grouped convolution layer GConv. As shown in Figure 3, the average pooling layer AVGPool in the convolution basic block BasicBlock calculates the number of features corresponding to each channel output by the second grouped convolution layer GConv based on the number of features contained in the original image feature map received by its input layer. This is then passed through the fully connected layer Concat to obtain the features output by BasicBlock.

[0062] In this embodiment, the dilated convolution Dilated Conv is introduced into the grouped convolution layer GConv, which can obtain a larger receptive field (RF) in the time series dimension. The receptive field refers to the size of the area perceived by each output unit in the input data, that is, the size of the area perceived by the features output by the grouped convolution layer GConv (the features on the hidden layer in Figure 5) in the features of its input. Since the features input to the grouped convolution layer GConv are high-frequency waveform signals and long time series data, maintaining the receptive field of the neural network model is very important for accurate modeling. Therefore, it is crucial and necessary to consider the RF of the neural network model (Convolutional Neural Network, CNN). By replacing the convolution kernel in GConv with a dilated convolution with a smaller kernel size and a dilation rate, the parameters and computational cost can be reduced without reducing the receptive field. RF l =RFl-1 +(K l -1)×S l

[0063] Where RF l The receptive field size of the input feature received by the grouped convolution layer GCon in the lth BasicBlock, the lth from the left in Figure 2, RF l-1 is the receptive field size of the input feature received by the grouped convolution layer GCon in the l-1 BasicBlock, S l It is the convolution operation of the lth group convolution layer GCon.

[0064] In this embodiment, as shown in FIG3 , BasicBlock performs a shuffle operation after performing group convolution, and then performs group convolution again. Finally, each channel on the output layer of BasicBlock is composed of the features on its input layer, thereby improving the accuracy of the model in which BasicBlock is located.

[0065] The BasicBlock input layer, which is the grouped convolution layer GConv, has the following relationship: n, m, and g.

[0066] Among them, the value of m must be a composite number, m = a × b, where a is the number of groups covered by the hidden layer, and b is the number of channels in each group. Assuming that the channel information transmission amount of ordinary convolution is M times that of void convolution, then

[0067] In one embodiment, the group convolution layer GConv in each convolution basic block (BasicBlock) is used to divide the blood pressure features received by the GConv input layer into multiple groups and apply a different convolution kernel to each group, that is, a different dilated convolution (Dilated Conv) is applied to increase the width of the network and learn more features between physiological signal waveforms. GConv can reduce the total number of parameters and the total computational cost of the neural network model because each GConv involves a small number of parameters (Params) and a small number of computational FLOPs.

[0068] Where C in is the number of channels of input data, C out is the number of channels of the output data, g is the fractional number, K is the convolution kernel size, that is, the size of the hole convolution, L out is the sequence length of the output data.

[0069] In one embodiment, S200 includes the following specific steps S201 to S206:

[0070] S201 , calculating the difference between the arterial systolic pressure SBP and the arterial diastolic pressure DSP, and recording it as the systolic-diastolic pressure difference.

[0071] S202 , multiplying the systolic and diastolic pressure difference by a coefficient less than one and then adding the result to the arterial diastolic pressure to obtain the plateau arterial blood pressure.

[0072] S203, counting the duration for which the MAP is continuously lower than the set blood pressure value.

[0073] S204, screening out the trainees in the database whose duration is longer than the set duration and recording them as target trainees.

[0074] S205 , filtering out from the database the training arterial pressure original waveform and / or the training electrocardiogram signal original waveform and / or the training pulmonary artery pressure original waveform related to the blood pressure signal of the target trainee and the classification training label corresponding to the target trainee.

[0075] The database in this embodiment comes from multiple existing data sets (MIMIC-III, Vital-DB). The database records the systolic blood pressure (SBP) and diastolic blood pressure (DSP) of multiple trainees, as well as the original waveforms of training arterial pressure, training ECG signals, and training pulmonary artery pressure. It also records whether each trainee has AHE. If AHE occurs, the trainee is marked as positive, that is, the classification training label is 1; otherwise, the trainee is marked as negative, that is, the classification training label is 0.

[0076] As shown in Figure 7 , AHE is defined as when MAP drops below 65 mmHg and persists for at least 5 minutes, with 65 mmHg as the set blood pressure value and 5 minutes as the set duration.

[0077] Where DBP and SBP can be extracted from the ABP waveform.

[0078] S206 , using the training arterial pressure original waveform and / or the training electrocardiogram signal original waveform and / or the training pulmonary artery pressure original waveform and the classification training label as the training data set.

[0079] Step S300 trains the neural network model, that is, inputting the training arterial pressure original waveform and / or the training electrocardiogram signal original waveform and / or the training pulmonary artery pressure original waveform into the neural network model. As shown in FIG8 , the above three original waveforms within ten minutes can be input into the neural network model to predict whether the trainee will experience AHE phenomenon after 5 minutes. The AHE label is input to the neural network, and the label is compared with the classification training label. If the two are inconsistent, the parameters of the neural network model are adjusted until the two are consistent to complete the training of the model.

[0080] In one embodiment, a method for classifying physiological signal waveforms is provided. This method is implemented using the aforementioned waveform classification model. The trained model is used to predict the probability of AHE, as shown in FIG9 . Specifically, the waveform classification model is used to predict the probability of AHE. Specifically, the three waveforms (ABP, PAP, and ECG) are input into the model. The model predicts the probability of AHE and converts the probability into two labels (0 and 1) for output. If an AHE event is predicted, the device hosting the waveform classification model issues an alarm so that necessary measures can be taken promptly in a clinical setting. This embodiment includes the following steps:

[0081] S400: extracting an original feature graph of a physiological signal waveform related to a blood pressure signal.

[0082] S500, applying a convolution module to the original feature map to obtain a final feature map, wherein the convolution module includes a plurality of convolution basic blocks cascaded in sequence, each of the convolution basic blocks includes a first grouped convolution layer, a shuffle operation layer, a second grouped convolution layer, a fully connected layer, and an output layer cascaded in sequence, wherein the convolution kernel of the first grouped convolution layer is a dilated convolution.

[0083] S600: Classify the physiological signal waveform according to the final feature map to obtain a classification result.

[0084] S400 to S600 obtain an original feature map related to the blood pressure signal based on multiple physiological signal waveforms of the predicted person, and then further process the original feature map using a convolution module to obtain a final feature map. According to the final feature map, the multiple physiological signal waveforms of the predicted person are classified to obtain a classification result. This classification result is used to characterize whether the predicted person will experience an acute hypotension event AHE.

[0085] In one embodiment, step S400 includes the following specific steps:

[0086] S401 , determining an original waveform of arterial pressure and / or an original waveform of electrocardiogram signal and / or an original waveform of pulmonary arterial pressure related to the blood pressure signal.

[0087] The original waveform of arterial pressure records the corresponding relationship between the original arterial pressure (ABP) of the predicted person and time. The original waveform of electrocardiogram (ECG) records the corresponding relationship between the original electrocardiogram (ECG) of the predicted person and time. The original waveform of pulmonary artery pressure records the corresponding relationship between the original pulmonary artery pressure (PAP) of the predicted person and time. ABP, ECG, and PAP are derived from the physiological signals of the predicted person collected by the ICU bedside vital sign monitor.

[0088] S402: Use the original waveform of arterial pressure and / or the original waveform of electrocardiogram signal and / or the original waveform of pulmonary arterial pressure as the physiological signal waveform.

[0089] S403: Apply a feature extraction module to the physiological signal waveform to obtain an original feature map, wherein the feature extraction module includes a 1x1 convolution layer Convld, a normalization layer BatchNom, and an activation function layer ReLU that are cascaded in sequence.

[0090] In this embodiment, the physiological signal waveform includes the original waveform of arterial pressure, the original waveform of the electrocardiogram signal, and the original waveform of pulmonary artery pressure. These three waveforms are simultaneously input into the feature extraction module, which cross-processes the three waveforms to extract the original feature map related to blood pressure.

[0091] In one embodiment, the specific process of step S600 is as follows: applying a classification module to the final feature map to classify the physiological signal waveform to obtain a classification result, wherein the classification module includes a cascaded average pooling layer AVG Pool and a classifier FC Classifier.

[0092] The so-called classification is to classify the predicted person who provides the physiological signal waveform, and the classification is used to characterize whether the predicted person will suffer from an acute hypotensive event AHE.

[0093] The following comparative experiments demonstrate the effectiveness of the neural network model proposed in this invention:

[0094] The existing model is ResNet, and its parameters are adjusted to determine the optimal number of channels, kernel size and stride. The neural network of the present invention sets 8 basic blocks.

[0095] The task was to use 10 minutes of ABP, PAP, and ECG waveforms from the MIMIC-III and Vital-DB datasets to predict whether a patient would experience AHE five minutes later. The results of the final comparative experiment showed that the proposed model achieved an accuracy of 90.18% and 88.13% on the MIMIC-III and Vital-DB datasets, respectively, achieving the highest accuracy among all models. Furthermore, the proposed model exhibited minimal parameter count, computational complexity, CPU memory usage, and GPU memory usage.

[0096] To simulate a clinical environment with limited computing resources, an NVIDIA GeForce GTX 1080 graphics card was used for the experiment. By adjusting parameters, the model of the present invention was used to predict whether the patient would experience AHE after 5 minutes, 10 minutes, and 15 minutes on the MIMIC-III and Vital-DB datasets using a variety of original signal waveforms. On both datasets, the model achieved the best accuracy: MIMIC-III dataset: 5 minutes: 92.04%, 10 minutes: 89.38%, 20 minutes: 87.55%; Vital-DB dataset: 5 minutes: 90.77%, 10 minutes: 88.89%, 20 minutes: 85.38%

[0097] In summary, using multiple raw waveform signals as model input avoids the tedious and complex feature extraction process. This greatly reduces the impact of inaccurate feature extraction due to individual patient differences, reduces system complexity, and improves model accuracy.

[0098] The lightweight model designed in this paper is suitable for use in ICU bedside vital signs monitoring environments with limited computing resources, and can provide highly accurate, real-time predictions of acute hypotensive events. This means that in real-world clinical applications, regardless of the computing power or storage space, rapid and effective monitoring and prediction can be performed, allowing timely implementation of necessary measures. This also provides important support for patient safety and treatment.

[0099] This embodiment also provides a waveform classification model training device, as shown in FIG10 , which includes the following components:

[0100] Model construction module 01 is used to construct a neural network model including a convolution module, wherein the convolution module includes a plurality of convolution basic blocks cascaded in sequence, each of which includes a first group convolution layer, a shuffle operation layer, a second group convolution layer, a fully connected layer, and an output layer cascaded in sequence, wherein the convolution kernel of the first group convolution layer is a dilated convolution;

[0101] Data screening module 02, for screening out a training data set from a database based on arterial diastolic pressure and arterial systolic pressure, wherein the database is used to store training waveforms of physiological signals related to blood pressure signals;

[0102] The training module 03 is used to train the neural network model using the training data set to obtain a waveform classification model.

[0103] In another embodiment, a physiological signal waveform classification device is provided, comprising:

[0104] The original feature map extraction module is used to extract the original feature map of the physiological signal waveform related to the blood pressure signal.

[0105] A feature optimization module is used to apply a convolution module to the original feature map to obtain a final feature map, wherein the convolution module includes a plurality of convolution basic blocks cascaded in sequence, each of the convolution basic blocks includes a first group convolution layer, a shuffle operation layer, a second group convolution layer and a fully connected layer and an output layer cascaded in sequence, wherein the convolution kernel of the first group convolution layer is a dilated convolution.

[0106] The result classification module is used to classify the physiological signal waveform according to the final feature map to obtain a classification result.

[0107] Original feature map extraction module, including:

[0108] a waveform analysis unit, configured to determine an original waveform of arterial pressure and / or an original waveform of electrocardiogram signal and / or an original waveform of pulmonary artery pressure related to the blood pressure signal;

[0109] a waveform selection unit, configured to use the original waveform of arterial pressure and / or the original waveform of electrocardiogram signal and / or the original waveform of pulmonary artery pressure as the physiological signal waveform;

[0110] A feature extraction unit is used to apply a feature extraction module to the physiological signal waveform to obtain an original feature map, wherein the feature extraction module includes a 1x1 convolution layer, a normalization layer, and an activation function layer cascaded in sequence.

[0111] Based on the above embodiments, the present invention also provides a terminal device, whose principle block diagram can be shown in Figure 11. The terminal device includes a processor, a memory, a network interface, and a display screen connected via a system bus. Among them, the processor of the terminal device is used to provide computing and control capabilities. The memory of the terminal device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the terminal device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a physiological signal waveform classification method is implemented. The display screen of the terminal device can be a liquid crystal display or an electronic ink display.

[0112] Those skilled in the art will understand that the principle block diagram shown in Figure 11 is only a block diagram of a partial structure related to the solution of the present invention, and does not constitute a limitation on the terminal device to which the solution of the present invention is applied. The specific terminal device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0113] In one embodiment, a terminal device is provided. The terminal device includes a memory, a processor, and a physiological signal waveform classification program stored in the memory and executable on the processor. When the processor executes the physiological signal waveform classification program, the following operating instructions are implemented:

[0114] Extracting original feature maps of physiological signal waveforms related to blood pressure signals;

[0115] Applying a convolution module to the original feature map to obtain a final feature map, wherein the convolution module includes a plurality of convolution basic blocks cascaded in sequence, each of the convolution basic blocks includes a first grouped convolution layer, a shuffle operation layer, a second grouped convolution layer, a fully connected layer, and an output layer cascaded in sequence, wherein the convolution kernel of the first grouped convolution layer is a dilated convolution;

[0116] The physiological signal waveform is classified according to the final feature map to obtain a classification result.

[0117] Those skilled in the art will appreciate that all or part of the processes in the above-described embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described embodiments. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A training method for a waveform classification model, which is applied to the classification of physiological signal waveforms. Characterized in that, It includes: Construct a neural network model including a convolutional module, where the convolutional module includes a number of cascaded convolutional basic blocks in sequence. Each convolutional basic block includes a first grouped convolutional layer, a shuffling operation layer, a second grouped convolutional layer, a fully connected layer, and an output layer cascaded in sequence, where the convolutional kernel of the first grouped convolutional layer is a dilated convolution; According to the diastolic arterial blood pressure and systolic arterial blood pressure, screen out a training data set from the database, where the database is used to store physiological signal training waveforms related to blood pressure signals; Use the training data set to train the neural network model to obtain a waveform classification model.

2. The training method for a waveform classification model according to claim 1, Characterized in that, The step of screening out a training data set from the database according to the diastolic arterial blood pressure and systolic arterial blood pressure, where the database is used to store physiological signal training waveforms related to blood pressure signals, includes: Calculate the difference between the systolic arterial blood pressure and the diastolic arterial blood pressure, denoted as the systolic-diastolic blood pressure difference; Multiply the systolic-diastolic blood pressure difference by a coefficient less than one and then add the diastolic arterial blood pressure to obtain the mean arterial blood pressure; According to the mean arterial blood pressure, screen out a training data set from the database, where the database is used to store physiological signal training waveforms related to blood pressure signals.

3. The training method for a waveform classification model according to claim 2, Characterized in that, The step of screening out a training data set from the database according to the mean arterial blood pressure, where the database is used to store physiological signal training waveforms related to blood pressure signals, includes: Statistical the duration corresponding to the mean arterial blood pressure continuously being lower than the set blood pressure value; Screen out the trainers corresponding to the duration greater than the set duration from each trainer in the database, denoted as target trainers; Screen out the original training arterial pressure waveforms and / or original training electrocardiogram signal waveforms and / or original training pulmonary artery pressure waveforms related to the blood pressure signals of the target trainers from the database, as well as the classification training labels corresponding to the target trainers; Use the original training arterial pressure waveforms and / or the original training electrocardiogram signal waveforms and / or the original training pulmonary artery pressure waveforms, and the classification training labels as the training data set.

4. A method for classifying physiological signal waveforms, Characterized in that, Apply the waveform classification model trained by the waveform classification model training method according to any one of claims 1-3. The method for classifying physiological signal waveforms includes: Extract the original feature map of the physiological signal waveform related to the blood pressure signal; Apply the convolutional module included in the waveform classification model to the original feature map to obtain a final feature map, where the convolutional module includes a number of cascaded convolutional basic blocks in sequence. Each convolutional basic block includes a first grouped convolutional layer, a shuffling operation layer, a second grouped convolutional layer, a fully connected layer, and an output layer cascaded in sequence, where the convolutional kernel of the first grouped convolutional layer is a dilated convolution; Classify the physiological signal waveform according to the final feature map to obtain a classification result.

5. The physiological signal waveform classification method according to claim 4, wherein, the convolutional basic block further includes an average pooling layer, and the average pooling layer is cascaded with the fully connected layer.

6. The physiological signal waveform classification method according to claim 4, wherein, extracting the original feature map of the physiological signal waveform related to the blood pressure signal includes: determining the original arterial pressure waveform and / or the original electrocardiogram signal waveform and / or the original pulmonary artery pressure waveform related to the blood pressure signal; using the original arterial pressure waveform and / or the original electrocardiogram signal waveform and / or the original pulmonary artery pressure waveform as the physiological signal waveform; applying a feature extraction module to the physiological signal waveform to obtain an original feature map, and the feature extraction module includes a 1x1 convolutional layer, a normalization layer, and an activation function layer cascaded in sequence.

7. The physiological signal waveform classification method according to claim 4, wherein, classifying the physiological signal waveform according to the final feature map to obtain a classification result includes: applying a classification module to the final feature map to classify the physiological signal waveform to obtain a classification result, and the classification module includes an average pooling layer and a classifier cascaded.

8. A training device for a waveform classification model, wherein, the device includes: a model construction module for constructing a neural network model including a convolutional module, the convolutional module includes a plurality of convolutional basic blocks cascaded in sequence, and each convolutional basic block includes a first grouped convolutional layer, a shuffling operation layer, a second grouped convolutional layer, a fully connected layer, and an output layer cascaded in sequence, wherein the convolutional kernel of the first grouped convolutional layer is a dilated convolution; a data screening module for screening a training data set from a database according to diastolic arterial blood pressure and systolic arterial blood pressure, and the database is used to store physiological signal training waveforms related to blood pressure signals; a training module for training the neural network model using the training data set to obtain a waveform classification model.

9. A terminal device, wherein, the terminal device includes a memory, a processor, and a training program for a waveform classification model stored in the memory and executable on the processor. When the processor executes the training program for the waveform classification model, the steps of the waveform classification model training method according to any one of claims 1-3 are implemented; alternatively, the terminal device includes a memory, a processor, and a physiological signal waveform classification program stored in the memory and executable on the processor. When the processor executes the physiological signal waveform classification program, the steps of the physiological signal waveform classification method according to any one of claims 4-7 are implemented.

10. A computer-readable storage medium, wherein, a training program for a waveform classification model is stored on the computer-readable storage medium. When the training program for the waveform classification model is executed by a processor, the steps of the waveform classification model training method according to any one of claims 1-3 are implemented; Alternatively, a physiological signal waveform classification program is stored on the computer-readable storage medium. When the physiological signal waveform classification program is executed by a processor, the steps of the physiological signal waveform classification method according to any one of claims 4-7 are implemented.

Images