TW202343477A - Method for osa severity detection using recording-based electrocardiography signal - Google Patents

Abstract

The present invention provides a method for OSA (Obstructive Sleep Apnea) severity detection using recording-based electrocardiography (ECG) Signal. The major feature of the present invention emphasizes on using a recording-based ECG Signal as an input, which is different from the deep learning-based prior art of using segment-based signals as an input to a model, and the segment-based signals has only two classification results, i.e. normal or apnea. The present invention provides a method for a model to detect and output directly a value of apnea-hypopnea index (AHI) for the OSA Severity.

Description

Method for detecting obstructive sleep apnea symptoms using whole electrocardiogram signals

The present invention relates to a method for detecting obstructive sleep apnea symptoms, in particular to using a recording-based electrocardiogram (ECG) signal as input to directly detect and output obstructive sleep apnea symptoms (OSA). ) Numerical result of the severity of apnea or hypopnea index (apnea-hypopnea index, AHI).

Please see Figure 1. All current deep learning technologies use the recognition of segment-based signals as model input, and only provide two types of recognition models for obstructive sleep apnea (OSA), namely Normal (normal) or abnormal (apnea). An electrocardiogram signal (ECG signal) 1 is cut into K ECG slice signals 2, which are respectively input to two types of identification models 3 to perform identification, and normal or abnormal identification results 4 are obtained. Then the OSA severity is calculated 5 and the identification result is displayed 6 .

AHI<15中度：15

AHI<30重度：AHI

30由上例可得AHI=25，屬於中度患者。 For example, calculate the OSA severity in Figure 1: Assume that a person sleeps for 8 hours. If you want to identify the OSA severity, use the entire 8-hour sleep period for detection. However, it is not a fixed 8 hours to detect. There is no limit to the length of sleep time. 8 hours is just an example. Assume: The total time length of a signal is equal to 8 hours; the time length of a slice signal: T = 60 seconds; a total of K = 480 slice signals are cut out (the total time length of a signal is equal to K * T = 28800 seconds = 8 hours ); Assume that a total of L=200 slice signals are identified through the model as abnormal; therefore, the OSA severity index apnea-hypopnea index (AHI) can be calculated: AHI=L/(K * T/3600), this equation means Refers to the average number of abnormalities per hour. Normal: AHI<5. Mild: 5

AHI<15 Moderate: 15

AHI<30 Severe: AHI

30 From the above example, we can get AHI=25, which is a moderate patient.

There are three disadvantages of this method: First, the identification process is quite complicated. The second is that the accuracy of the recognition model will be affected by different slice signal time lengths. The third is the sliced signal data set used in the training process of the recognition model. The labeling work is very complicated and costs considerable manpower and time.

The purpose of the present invention is to provide a method for detecting obstructive sleep apnea (OSA) symptoms using the entire electrocardiogram signal as input. The content of the present invention is described as follows.

First, an obstructive sleep apnea (OSA) severity detection model is established.

An electrocardiogram signal (ECG) public data set is used as training data and input into the OSA severity detection model for training to complete a model.

Input an entire electrocardiogram signal (ECG) into the model and directly display the AHI value and the corresponding OSA severity classification result (i.e. normal, mild, moderate or severe).

The entire electrocardiogram signal (ECG) is input into the model, and is processed by a feature map extraction layer based on a convolutional neural network, a global average pooling layer, a fully connected layer, and an output layer to obtain the AHI value. and the corresponding four-category results of OSA severity.

1:ECG signal

2:K ECG slice signals

3: Two types of OSA identification models

4: Identification results of K slice signals

5: Calculation of OSA severity

6: Display the identification results

21: OSA severity detection model

22: Display detection results

311, 312, 313, 321, 323, 324, 325, 331, 332, 334, 335: convolutional layer

314: Pooling layer

322, 333: Addition

340: Global average pooling layer

350: Fully connected layer

360:Output layer

41:ECG public data set

42: Establish OSA severity detection model

43: Using ECG training set

44:Training model

45: Is it convergent?

46: Model training completed

51: Subject

52: Wearable devices

53: Display diagnosis results

Figure 1 is a schematic diagram of two types of recognition models for obstructive sleep apnea (OSA) using segment-based signal recognition as model input.

Figure 2 is a schematic diagram of the present invention using the entire electrocardiogram signal (ECG signal) as input to directly detect and output the AHI value of obstructive sleep apnea symptoms (OSA) and the corresponding severity classification results.

Figure 3 details an embodiment of the OSA severity detection model 21 architecture of Figure 2.

Figure 4 is a schematic diagram of how the present invention trains and generates an OSA severity detection model.

FIG. 5 is a schematic diagram of a subject carrying a wearable device for measuring ECG signals while performing OSA diagnosis according to the present invention.

Figure 2 illustrates that the present invention uses the entire electrocardiogram signal (ECG signal) 1 as input to directly detect and output the AHI value of obstructive sleep apnea (OSA) and the corresponding four severity classification results 22.

The present invention provides the OSA severity detection model 21 to directly output the AHI value and the corresponding OSA severity result, that is, four categories: normal, mild, moderate, and severe. Directly input the entire amount (for example, 8 hours, but no time limit) of electrocardiogram signals (ECG signals) 1 into the model for identification, and directly display the detection results 22 .

Figure 3 details an embodiment of the OSA severity detection model 21 architecture of Figure 2. The content of Figure 3 can be subdivided into four parts: feature map extraction layer 311~335 based on convolutional neural network, global average pooling layer 340, fully connected layer 350, and output layer 360. The input of this patented model is the entire ECG signal 1. There is no limit to the time length of the input signal and it can be a signal of any length. However, the input signal time length of the existing methods (two types of OSA identification models) is fixed.

Feature map extraction layer 311~335 based on convolutional neural network: The feature map extraction layer uses a convolutional neural network (CNN) to extract feature maps of the input signal. . Convolutional neural network is the most commonly used method for deep learning. The biggest feature of convolutional neural network is that it can automatically extract the characteristic information of the input signal through model training, called a feature map, and then use the feature map to perform identification tasks. This method can effectively improve the identification accuracy. The components of a convolutional neural network include a convolution layer, an activation function, and a pooling layer. Through repeated multi-level series or parallel connections of these components, a convolutional neural network can be constructed. A wide variety of convolutional neural networks with different architectures. In this embodiment, the convolution layers 311 to 313 and the pooling layer 314 are the first-level feature map extraction; the convolution layer 321, the addition 322, the convolution layer 323, the convolution layer 324, and the convolution layer 325 are the second-level features. Image extraction; and convolution layer 331, convolution layer 332, addition 333, convolution layer 334, and convolution layer 335 are the third level of feature map extraction.

Global average pooling layer 340: Use global average pooling pooling) method, which calculates the average of each feature map as the output of this pooling layer. Using this method, input signals of different lengths can be converted into outputs of the same length. In other words, in this way, the input signal of this patented model can be of any length.

Fully connected layer 350: integrates the highly abstract features obtained previously and then passes them to the output layer 360.

Output layer 360: Use the Rectified Linear Unit (ReLU) activation function to output a value greater than or equal to zero, which is the AHI value.

In this embodiment, all convolutional layers use one-dimensional (1-dimension) convolution kernels. The convolutional layer 311 uses 32 convolution kernels of size 20, which are represented here by (32,20). The convolution layer 312 uses the convolution kernel of (64,20); the convolution layer 313 uses the convolution kernel of (128,5); the convolution layer 321 uses the convolution kernel of (128,3); the convolution layer 323 uses (128,3) ) convolution kernel; convolution layer 324 uses the convolution kernel of (64,1); convolution layer 325 uses the convolution kernel of (128,3); convolution layer 331 uses the convolution kernel of (128,3); convolution layer 332 uses the convolution kernel of (128,3); the convolution layer 334 uses the convolution kernel of (128,3); the convolution layer 335 uses the convolution kernel of (64,1).

In this embodiment, if the input ECG signal 1 is input data of 2,160,000 sampling points sampled at a time length of 6 hours and a sampling frequency of 100Hz, the convolution operation of the convolution layer 311 will be converted into 32 features with a size of 108,000 Figure, this feature map is a two-dimensional (2-dimension) array, represented by 108,000×32. Next, the convolution layer 312 performs a convolution operation on the feature map output by the convolution layer 311 to generate a 5,400×64 feature map. The feature map is then subjected to a convolution operation through the convolution layer 313 to obtain a 1,080×128 feature map. The pooling layer 314 uses a sliding window of size 2 to apply a maximum pooling method (max pooling) to the feature map output by the convolution layer 313 to obtain a 540×128 feature map.

Next, the convolution layer 321 performs a convolution operation on the feature map output by the pooling layer 314 to generate a 540×128 feature map, and then the convolution layer 323 performs a convolution operation to obtain a 540×128 feature map. Then, the feature map output by the pooling layer 314 and the feature map output by the convolution layer 323 will be added in an addition 322 to obtain a fused 540×128 feature map. The convolution layer 324 performs a convolution operation on the feature map output by the addition 322 to obtain a 540×64 feature map. The convolution layer 325 performs a convolution operation on the feature map output by the convolution layer 324 to obtain a 540×128 feature map.

Then, the convolution layer 331 performs a convolution operation on the feature map output by the convolution layer 325 to generate a 540×128 feature map, and then performs a convolution operation through the convolution layer 332 and the convolution layer 334 to maintain the same dimension of 540×128 features. Figure. Then, the feature map output by the convolution layer 325 and the feature map output by the convolution layer 334 are added in an addition 333 to obtain a fused 540×128 feature map. Finally, the convolution layer 335 performs a convolution operation on the feature map output by the addition 333 to obtain a 540×64 feature map.

This embodiment then uses the global average pooling method on the feature map output by the convolution layer 335 in the global average pooling layer 340 to obtain the average value of 64 feature maps. It is worth mentioning that using the global average pooling layer can convert input data of different lengths into outputs of the same length. In this way, the input signal of this patented model can be of any length. Then, the features output by the global average pooling layer 340 will be connected to the fully connected layer 350 with 16 neurons, and the fully connected layer 350 will then be connected to the output layer 360 with only 1 neuron. The output layer 360 uses the ReLU activation function to output a value greater than or equal to zero, which is the AHI value. Finally, the display detection result 22 will display the AHI value and the OSA severity corresponding to this value, that is, four categories: normal, mild, moderate, and severe.

Figure 4 illustrates how the present invention trains and generates a model. First obtain a published ECG Data set 41 is used to establish an OSA severity detection model 42 , and the ECG training set 43 is used to input the OSA severity detection model for training 44 . During the training process, it will be checked whether the model has converged45. If the model has reached convergence, the training will end, otherwise the training will continue. The trained model 46 can be used to perform OSA diagnosis.

Nowadays, wearable devices or portable devices capable of measuring ECG signals are becoming more and more popular. It is very convenient for OSA diagnosis by analyzing ECG signals, allowing users to self-test at home. Please see Figure 5, which shows that the subject 51 is wearing a wearable device 52 for measuring ECG signals, and the ECG signals are obtained and input into the OSA severity detection model 21 established by the present invention, OSA diagnosis is performed, and the AHI value is directly displayed. Corresponding OSA severity: normal, mild, moderate or severe diagnosis results53.

The spirit and scope of the present invention are determined by the following patent application scope and are not limited to the above embodiments.

51: Subject

52: Wearable devices

1:ECG signal

21: OSA severity detection model

53: Display diagnosis results

Claims (3)

A method of detecting obstructive sleep apnea symptoms using the entire electrocardiogram signal, including: 1. Establish a severity detection model of obstructive sleep apnea (OSA); 2. Use an electrocardiogram signal (ECG) public data set as training data, input it into the OSA severity detection model for training, and complete a model; 3. Input a whole set of electrocardiogram signals (ECG) into the model to directly display the apnea or hypopnea index (AHI) value and the OSA severity classification result corresponding to the AHI value (i.e. normal, mild). severe, moderate or severe). For example, the method of claim 1 using the entire electrocardiogram signal to detect obstructive sleep apnea symptoms, wherein the entire electrocardiogram signal (ECG) is input into the model and undergoes a feature map extraction layer based on a convolutional neural network, a global The average pooling layer, a fully connected layer, and an output layer are processed to obtain the AHI value and the corresponding four classification results of OSA severity. For example, the method of claim 1 for detecting obstructive sleep apnea symptoms using a whole electrocardiogram signal, wherein the whole electrocardiogram signal (ECG) is obtained by an ECG wearable device.

Images