CN111166294B - Automatic sleep apnea detection method and device based on inter-heartbeat period - Google Patents

Abstract

The present invention provides an automatic detection method and device for sleep apnea based on heartbeat interval, which are used to solve the problems of inaccurate and low-precision detection of sleep apnea in the prior art. The method for automatic detection of sleep apnea first collects heartbeat interval information during sleep, then automatically extracts features from the heartbeat interval information through a residual neural network, further extracts heart rate variability features, and then extracts the automatically extracted features. It is fused with the heart rate variability feature set to determine whether apnea occurs. The present invention combines the heart rate variability feature by extracting the heartbeat interval feature in the ECG signal of the electrocardiogram, and deeply analyzing the feature in the residual neural network, and all the weights in the residual neural network can be used in clinical practice. The fine-tuning on the above improves the spirituality, accuracy and precision of sleep detection; at the same time, only single-lead ECG information is required, and the collection process is simple and convenient, and has considerable universality.

Description

A method and device for automatic detection of sleep apnea based on heartbeat interval

technical field

The invention belongs to the field of medical data monitoring, in particular to a method and device for automatic detection of sleep apnea based on heartbeat interval.

Background technique

Sleep Apnea (SA) is a common breathing disorder that causes sleep disorders. During REM sleep, complete obstruction of the airway by the genioglossus muscle, tendon, and adipose tissue for at least 10 seconds, or hypopnea with a decrease in airflow of more than 50% and oxygen desaturation by more than 3% for 10 seconds, is called for apnea. According to the survey, sleep apnea has been affecting people all over the world for more than 30 years, and in 2008, the number of patients had reached 6% of global adults. An increasing number of patients with apnea are at risk of cardiovascular disease and clinical depression due to undiagnosed and uninterrupted interventions. Numerous experiments have investigated the hazards of apnea, and the results have shown a strong correlation between apnea and several endogenous physiological phenomena and diseases.

Sleep apnea is long-term, harmful, but can be treated after diagnosis. Treatments such as positive airway pressure (PAP) and palatopharyngoplasty (PPP) are very effective in early diagnosis. Therefore, timely diagnosis is crucial for the treatment process of apnea. Polysomnogram (PSG) signals have traditionally been used to diagnose apnea syndrome severity, but when PSG signals are collected to detect apnea, patients are required to sleep overnight with invasive devices, medical specialists and clinics surroundings. This traditional apnea detection process is too complex and expensive, requiring new methods that can be performed quickly and accurately on non-invasive devices.

In the prior art, an effective solution for detecting sleep apnea is to use a single-lead electrocardiogram (ECG) signal to detect apnea, and currently, various functions are usually designed manually. For example, by extracting features from ECG signals, put them into various classifiers such as proximity algorithm (kNN) and support vector machine (SVM) for classification, or perform automatic feature extraction and classification through deep neural networks. But the ECG signal sampled at around 100Hz limits the depth of the neural network, which affects the accuracy and precision of sleep breathing detection. Therefore, the existing sleep breathing detection methods have certain defects in terms of user experience, flexibility and accuracy.

SUMMARY OF THE INVENTION

The embodiments of the present invention provide a method and device for automatic detection of sleep apnea based on heartbeat interval. By extracting the heartbeat interval feature in the ECG signal of the electrocardiogram, and deeply analyzing the feature in the residual neural network, combined with the heart rate Variability characteristics, determine whether apnea occurs, and improve the accuracy and flexibility of sleep apnea detection.

In order to achieve the above purpose, the technical solutions adopted in the embodiments of the present invention are as follows.

In a first aspect, an embodiment of the present invention provides an automatic detection method for sleep apnea based on heartbeat interval, the method comprising the following steps:

Step S1, collects the human electrocardiogram ECG signal during sleep, and extracts the heartbeat interval information of adjacent electrocardiograms according to the R wave in the five waves of the electrocardiogram cycle PQRST, and generates a heartbeat interval time series;

Step S2, automatic feature extraction is performed on the heartbeat interval time series based on a preset residual neural network to obtain a first HRV feature;

Step S3, performing heart rate variability feature extraction based on the heartbeat interval time series to obtain a second heart rate variability feature;

Step S4, the first HRV feature automatically extracted by the residual neural network and the second HRV feature set are feature-fused, and input to the classifier together to determine whether apnea occurs.

Optionally, in the step S1, a heartbeat interval time series is generated, and further, a linear interpolation is used to resample the extracted heartbeat interval information at a frequency of 2 Hz to generate a heartbeat interval time series.

Optionally, in the step S2, automatic feature extraction is performed on the heartbeat interval time series, further comprising: using a convolutional neural network CNN and a back-propagation BP algorithm to train the residual neural network to complete automatic feature extraction.

Optionally, the convolutional neural network is composed of 33 one-dimensional convolutional layers in total, and 16 residual blocks are formed by combining with the structure of the batch normalization layer, the dropout layer, and the ReLu function layer.

Optionally, in the ReLu function layer, the one-dimensional convolutional layer is the core of the CNN, the input vector of the activation function h is X, the output is the input of the next layer, and the output of the first convolutional layer is Y ^(l) _conv is:

是卷积计算，B为偏置矩阵，W为具有固定大小的权值向量。In formula (1),

is the convolution calculation, B is the bias matrix, and W is the weight vector with fixed size.

Optionally, an average pooling layer is used between the convolutional layers.

Optionally, among the 16 residual blocks, a short-circuit connection is established, the input of the residual block is added to the output of the residual block, and a projection is performed in the short-circuit connection, so that the input size is the same as the output size. Matching, so as to complete the automatic extraction of neural network features.

Optionally, described step S3 heart rate variability characteristics, including: the mean value of heartbeat interval, the standard deviation of heartbeat interval, the skewness of heartbeat interval, the kurtosis of heartbeat interval, NN50 measurement value, pNN50 index , the standard deviation of the difference between adjacent heartbeat intervals, the square root of the square mean of the difference between adjacent heartbeat intervals, the Allan factor, the very low frequency of the heartbeat interval, the low frequency of the heartbeat interval, the high frequency of the heartbeat interval, The ratio of low frequency LF to high frequency HF in the heartbeat interval.

Optionally, judging whether apnea occurs in the step S4 is further:

等于1；当判别值x为负时，伪概率小于50％，判断当前时间段内患者未出现睡眠呼吸暂停，预测标签值

等于0。The first HRV feature automatically extracted by the residual neural network and the manually extracted second HRV feature are fused into a feature vector, and input into a layer of fully connected neural network classifier; the feature After the vector passes through the classifier, an unstandardized discriminant value x is obtained, and the discriminant value is converted into the pseudo probability of sleep apnea in the patient in the current time period through the output function sigmoid(x); when the discriminant value x is non-negative , the pseudo probability is greater than or equal to 50%, it is judged that the patient has sleep apnea in the current time period, and the label value is predicted

Equal to 1; when the discriminant value x is negative, the pseudo probability is less than 50%, it is judged that the patient does not have sleep apnea in the current time period, and the label value is predicted

equal to 0.

In a second aspect, an embodiment of the present invention provides an automatic detection device for sleep apnea based on heartbeat interval. The automatic detection device for sleep apnea includes: a heartbeat interval information collection module, a depth feature extraction module, and a heart rate variability feature. Extraction module, sleep apnea judgment module; wherein,

The heartbeat interval information collection module is connected with the depth feature extraction module and the heart rate variability feature extraction module at the same time, and is used to collect the ECG signal of the human body electrocardiogram during sleep. Extracting heartbeat interval information of the electrocardiogram, generating a heartbeat interval time series, and sending the heartbeat interval time series to the depth feature extraction module and the heart rate variability feature extraction module;

The depth feature extraction module is configured to automatically extract features according to the generated heartbeat interval time series and based on a preset residual neural network, and send the automatically extracted features to the sleep apnea judgment module;

The heart rate variability feature extraction module is configured to extract the heart rate variability feature according to the generated heartbeat interval time series, and send the heart rate variability feature to the sleep apnea judgment module;

The sleep apnea judging module is connected with the depth feature extraction module and the heart rate variability feature extraction module at the same time, and is used for judging whether sleep apnea occurs according to the received automatic extraction features and heart rate variability features.

The present invention has the following beneficial effects: the method and device for automatic detection of sleep apnea based on heartbeat interval according to the embodiment of the present invention first collect heartbeat interval information during sleep, and then use a residual neural network to analyze the heartbeat interval information. Automatically extract features, further extract heart rate variability features, and then fuse the automatically extracted features and heart rate variability feature sets to determine whether apnea occurs, which has great flexibility and improves the accuracy of sleep apnea detection. performance and detection accuracy.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

1 is a schematic flowchart of an automatic detection method for sleep apnea based on heartbeat interval according to an embodiment of the present invention;

2 is a flow chart of calculating a residual neural network in a method for automatic detection of sleep apnea according to an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of an apparatus for automatic detection of sleep apnea based on heartbeat interval according to an embodiment of the present invention.

Detailed ways

The technical problems, technical solutions and advantages of the present invention will be explained in detail below by referring to the exemplary embodiments. The exemplary embodiments described below are only for explaining the present invention, and should not be construed as limiting the present invention. It will be understood by those of ordinary skill in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and not in idealized or overly formal meanings unless defined herein to explain.

The embodiments of the present invention detect sleep apnea by extracting a heartbeat interval (HI) feature in an ECG signal, and performing an in-depth analysis of the feature in a residual neural network (Residual Network, RN). The residual neural network, combined with Heart Rate Variability (HRV), is a deep algorithm applied on wearable devices or smart phones by using network compression technology, and the residual neural network is a deep algorithm. All weights can be fine-tuned, giving a lot of flexibility compared to hand-designed features.

In order to facilitate the understanding of the embodiments of the present invention, the following will take several specific embodiments as examples for further explanation and description in conjunction with the accompanying drawings, and each embodiment does not constitute a limitation on the technical solutions of the present invention.

first embodiment

This embodiment provides an automatic detection method for sleep apnea based on a heartbeat interval, and FIG. 1 is a schematic flowchart of the automatic detection method for sleep apnea. As shown in Figure 1, the method for automatic detection of sleep apnea based on heartbeat interval includes the following steps:

In step S1, the ECG signal of the human body electrocardiogram during sleep is collected, and the heartbeat interval information of adjacent ECGs is extracted according to the R wave in the five waves of the electrocardiogram cycle PQRST to generate a heartbeat interval time series.

In this step, the collection of the human electrocardiogram ECG signal during sleep may be performed by using a simple and comfortable device such as a mobile device. Preferably, the sampling frequency is 100 Hz. The step of generating a heartbeat interval time series is further by using linear interpolation to resample the extracted heartbeat interval information at a frequency of 2 Hz to generate a heartbeat interval time series.

Step S2: Automatic feature extraction is performed on the heartbeat interval time series based on a preset residual neural network to obtain a first HRV feature.

In this step, the automatic feature extraction of the heartbeat interval time series is further: using a convolutional neural network (CNN) and a backpropagation (Backpropagation, BP) algorithm to train the residual neural network to complete the automatic feature extraction .

FIG. 2 is a flow chart of calculating the residual neural network in this embodiment. Through deep feature extraction, the first HRV feature is obtained. As shown in Figure 2, after the heartbeat interval time series with a duration of 3 minutes is input, 128 feature maps are first extracted through 5 one-dimensional convolutional layers of size 1*20, followed by batch normalization. The batch normalization (BN) layer is normalized and the ReLu function layer is activated, and then average pooled by an average pooling layer of size 1*2; followed by 16 identical residuals The blocks are connected in sequence, and each residual block is to add the incoming feature maps to itself after passing through 2 one-dimensional convolution layers of size 1*3, 2 BN layers, and 2 Relu function layers, short-circuiting The connection is used to add the input and output of the residual block; finally, through a convolutional layer of size 1*3 and an average pooling layer, the obtained 512 feature maps are the automatically extracted first heart rate variability feature.

The one-dimensional convolution layer is the core of CNN. In the convolution layer, the weight vector W with a fixed size will be multiplied by the input data one by one, which is the convolution operation. The convolution result plus the bias term B constitutes the activation Input to function h. The output of the activation function h will be the input to the next layer. The input vector is X, then the output Y of the lth convolutional layer is:

是卷积计算，B为偏置矩阵，W为具有固定大小的权值向量。In formula (1),

is the convolution calculation, B is the bias matrix, and W is the weight vector with fixed size.

In this embodiment, all activation functions h in the convolutional layer are ReLU functions to speed up training and reduce the final error rate in the neural network.

The ReLU function for input x is:

h _RelU (x)=max(0,x) (2)

There are two types of pooling layers in CNN: average pooling layers and max pooling layers. A max pooling layer used to reduce mean shift error caused by poor weight initialization, or using Xavier initialization in convolutional layers, can also reduce mean shift error. The average pooling layer can reduce the standard deviation error in the output of the convolutional layer. Noise caused by measurement errors and other physiological mechanisms can contribute to standard deviation errors in the heartbeat interval. In addition, an averaging filter can also be used for inter-beat correction. In this embodiment, an average pooling layer is used between the convolutional layers to reduce the standard deviation error of the heartbeat interval caused by measurement error and other physiological mechanisms in the output of the convolutional layer.

A batch normalization layer (BN layer) is used to reduce the internal covariate shift in the residual neural network, normalizing the standard deviation Var of the input X and the expectation E:

In Eq. (3), E(x) and Var(x) are the expectation and bias of the input X, x is the input sample, and ∈ is the smoothing term to avoid division by zero, which is usually set to be negligibly very small. positive number. The batch normalization layer (BN layer) in the neural network can realize the function of data whitening.

Among the 16 residual blocks of the residual neural network, each residual block is composed of two one-dimensional convolutional layers, two BN layers and two Relu function layers. The input X is used to define its output Y:

Y=F(X,{W}) (4)

In formula (4), F represents the mapping of the block, and the set {W} represents all parameters in the block, and the output Y _r of the residual block is obtained as:

Y _r =Y+X=F(X,{W})+X (5)

When the dimension of Y is greater than X, project X to keep the same dimension of both matrices.

Using the BP algorithm to calculate the gradient G _i of the _i -th parameter Wi is:

可能为零；当梯度等于零时，卷积层停止训练。When G _i is equal to zero,

May be zero; when the gradient equals zero, the convolutional layer stops training.

是Gradients in Residual Blocks

Yes

始终是残差网络中前一个残余块的输出，为非零。

Always the output of the previous residual block in the residual network, non-zero.

Among the 16 residual blocks, a short-circuit connection is established, the input of the residual block is added to its output, and a projection is made in the short-circuit connection so that the input size matches the output size.

Step S3, extracting the heart rate variability feature based on the heartbeat interval time series to obtain the second heart rate variability feature.

In this step, the heart rate variability feature is the small variation of heart rate between successive heartbeat cycles, and also refers to the small change between heartbeat intervals. The heart rate variability information can be obtained by extracting linear and nonlinear data features for the heartbeat interval. .

In the residual neural network calculation process shown in Figure 2, the heartbeat interval time series with a duration of 3 minutes is input, and artificial feature extraction is performed for each minute, so that the same feature can be calculated to obtain three results, and finally get The second HRV feature is shown in FIG. 2 , and the second HRV feature types include: the average value of the heartbeat interval, the standard deviation of the heartbeat interval, the skewness of the heartbeat interval, and the peak value of the heartbeat interval Degree, NN50 measure, pNN50 index, standard deviation of the difference between adjacent heartbeat intervals, square root of the mean squared difference between adjacent heartbeat intervals, Allan factor, very low frequency of heartbeat interval, low frequency of heartbeat interval , the high frequency of the heartbeat interval, the ratio of the low frequency LF to the high frequency HF of the heartbeat interval. All the second HRV features form a feature vector.

The second heart rate variability feature extraction based on the heartbeat interval time series further includes the following steps:

Step S301, calculating the mean of RR intervals (mean of RR intervals, Mean RR);

Step S302, calculating the standard deviation of the heartbeat interval (Standard deviation of RR intervals, SDRR);

Step S303, calculating the skewness of the heartbeat interval (Skewness of RR intervals, Skewness RR);

Step S304, calculating the kurtosis of the heartbeat interval (Kurtosis of RR intervals, Kurtosis RR);

Step S305, calculate the NN50 metric value (variant 1), and the NN50 metric value (variant 1) is the interval in which the duration of the first heartbeat interval in the adjacent heartbeat interval is greater than the duration of the second heartbeat interval by at least 50ms Number of pairs (number of pairs of adjacent normal to normal intervals differing by more than 50ms, NN50);

Step S306, calculate NN50 metric value (variant 2), described NN50 metric value (variant 2) is the number of adjacent heartbeat interval pairs that the second heartbeat interval exceeds the first heartbeat interval to exceed 50ms;

Step S307, calculating the pNN50 index of variant 1, dividing each NN50 index by the total number of heartbeat intervals (Percentof NN50 in the total number of RR intervals, PNN50);

Step S308, calculating the pNN50 index of variant 2, dividing each NN50 index by the total number of heartbeat intervals;

Step S309, calculating the standard deviation of the difference between adjacent heartbeat intervals (Standard deviation of Successive Difference between adjacent cycles, SDSD);

Step S310, calculating the square root of the mean square of difference between adjacent heartbeat intervals (The root mean square of difference between adjacent NN intervals, R MSSD);

Step S311, calculate the Allan factor A(T) for evaluation at time scales 5, 10, 15, 20 and 25s

In formula (8), N _i (T) is the number of QRS detection points in the window of length T, extending from iT to (i+1)T, and E is the expectation operator;

Step S312, calculating very low frequencies (VLF) in the heartbeat interval;

Step S313, calculating the low frequency (Low frequency, LF) of the heartbeat interval;

Step S314, calculating the high frequency (High Frequency, HF) of the heartbeat interval;

In step S315, the ratio of LF to HF in the heartbeat interval is calculated.

Step S4, the first HRV feature automatically extracted by the residual neural network and the second HRV feature set are feature-fused, and input to the classifier together to determine whether apnea occurs.

In this step, the first HRV feature automatically extracted by the residual neural network and the manually extracted second HRV feature are fused into a feature vector, which is input into the classifier. The feature vector obtains an unstandardized discriminant value x after passing through a layer of fully connected neural network classifier, and this discriminant value is converted into the pseudo probability of sleep apnea in the patient in the current period through the output function sigmoid(x). Between 0-100%, the sigmoid function of input x is represented as follows:

时，二元交叉熵损失函数为：When training, when the true label is y and the predicted label is

When , the binary cross-entropy loss function is:

的计算方法为：Predict labels based on pseudo-probabilities

The calculation method is:

等于1；当判别值x为负时，伪概率会小于50％，判断当前时间段内患者未出现睡眠呼吸暂停，预测标签值

等于0。When the discriminant value x is non-negative, the pseudo probability is greater than or equal to 50%, it is judged that the patient has sleep apnea in the current time period, and the label value is predicted.

Equal to 1; when the discriminant value x is negative, the pseudo probability will be less than 50%, judging that the patient does not have sleep apnea in the current time period, and predicting the label value

equal to 0.

It can be seen from the above technical solutions that the method for automatic detection of sleep apnea based on heartbeat interval described in this embodiment first collects heartbeat interval information during sleep, and then automatically extracts the heartbeat interval information through a residual neural network. The first HRV feature is extracted, the second HRV feature is further extracted, and the first HRV feature and the second HRV feature set are fused to determine whether apnea occurs. The method for automatic detection of sleep apnea only needs single-lead electrocardiogram information, the collection process is simple and convenient, and whether sleep apnea occurs can be accurately detected, thereby improving the spirituality, accuracy and precision of sleep detection.

Second Embodiment

This embodiment provides an automatic detection device for sleep apnea based on heartbeat interval, and FIG. 3 is a schematic structural diagram of the automatic detection device. As shown in Figure 3, the sleep apnea automatic detection device includes: a heartbeat interval information collection module, a depth feature extraction module, a heart rate variability feature extraction module, and a sleep apnea judgment module; wherein,

The heartbeat interval information collection module is connected with the depth feature extraction module and the heart rate variability feature extraction module at the same time, and is used to collect the human electrocardiogram ECG signal during sleep. Extracting the heartbeat interval information of adjacent ECGs, generating a heartbeat interval time series, and sending the heartbeat interval time series to the automatic feature extraction module and the heart rate variability feature extraction module;

The deep feature extraction module is configured to automatically extract features according to the generated heartbeat interval time series and based on a preset residual neural network to obtain a first heart rate variability feature, and send the automatically extracted features to the sleeper Apnea judgment module;

The heart rate variability feature extraction module is configured to perform heart rate variability feature extraction based on the heartbeat interval time series, obtain a second heart rate variability feature, and send the heart rate variability feature to the sleep apnea judgment module;

The sleep apnea judging module is connected with the depth feature extraction module and the heart rate variability feature extraction module at the same time, and is used for judging whether sleep apnea occurs according to the received automatic extraction features and heart rate variability features.

The device for automatic detection of sleep apnea based on heartbeat interval described in this embodiment is a technical solution corresponding to the automatic detection method for sleep apnea based on heartbeat interval described in the first embodiment, and the embodiments are the same or similar. With partial reference to each other, each embodiment focuses on the differences from the other embodiments. In particular, for the apparatus or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial descriptions of the method embodiments, which will not be repeated here.

It can be seen from the above technical solutions that the automatic detection device for sleep apnea based on the heartbeat interval described in this embodiment is suitable for a wide audience, has universality, and provides accurate and fast auxiliary disease diagnosis services for patients, which can be discovered in time. Endogenous risk of cardiovascular and cerebrovascular diseases.

The above descriptions are the preferred embodiments of the present invention. It should be noted that the present invention is not limited to the exemplary embodiments disclosed above, and the essence of the description is only to help those skilled in the relevant art to comprehensively understand the specific details of the present invention. For those of ordinary skill in the art, without departing from the principles of the present invention, several improvements and modifications, and easily conceivable changes or substitutions made within the technical scope of the present invention should be covered in the within the protection scope of the present invention.

Claims (6)

1. An automatic sleep apnea detecting device based on a heartbeat interval, comprising: the device comprises a heartbeat interval information acquisition module, a depth feature extraction module, a heart rate variability feature extraction module and a sleep apnea judgment module; wherein,

the heartbeat interval information acquisition module is simultaneously connected with the depth characteristic extraction module and the heart rate variability characteristic extraction module and is used for acquiring a human electrocardiogram ECG signal during sleeping, extracting heartbeat interval information of adjacent electrocardios by adopting 2Hz frequency according to R waves in five waves of an electrocardio period PQRST by using linear interpolation values according to the extracted heartbeat interval information, generating a heartbeat interval time sequence and sending the heartbeat interval time sequence to the depth characteristic extraction module and the heart rate variability characteristic extraction module;

the depth feature extraction module is used for training the residual error neural network by using a Convolutional Neural Network (CNN) and a Back Propagation (BP) algorithm according to the generated heartbeat interval time sequence and based on a preset residual error neural network, automatically extracting features to obtain a first heart rate variability feature, and sending the automatically extracted features to the sleep apnea judgment module; the convolutional neural network is composed of 33 layers of one-dimensional convolutional layers in total, and is combined with a batch normalization layer, a dropout layer and a ReLu function layer to form 16 residual blocks;

the heart rate variability feature extraction module is used for extracting heart rate variability features based on the heart rate interval time sequence, extracting artificial features for each minute in the heart rate interval time sequence with the duration of 3 minutes input in the residual error neural network to obtain a second heart rate variability feature, and sending the heart rate variability feature to the sleep apnea judgment module;

the sleep apnea judging module is simultaneously connected with the depth feature extracting module and the heart rate variability feature extracting module, and is used for performing feature fusion on the first heart rate variability feature and the second heart rate variability feature set, inputting the feature fusion into the classifier, and judging whether sleep apnea occurs.

2. The automatic sleep apnea detecting device of claim 1, wherein in the ReLu function layer, the one-dimensional convolutional layer is a core of CNN, an input vector of an activation function h is X, an output is an input of a next layer, and an output Y of the one-dimensional convolutional layer is:

in the formula (1), the reaction mixture is,

is a convolution calculationB is a bias matrix and W is a weight vector with a fixed size.

3. The automatic sleep apnea detection device of claim 1, wherein an averaging pooling layer is employed between said convolutional layers.

4. The automatic sleep apnea detecting device of claim 1, wherein a short circuit connection is established among said 16 residual blocks for adding the input of the residual block to the output of the residual block, and projection is performed in the short circuit connection to match the input size with the output size, so as to complete automatic extraction of neural network features.

5. An automatic sleep apnea detection device as recited in claim 1, wherein said heart rate variability feature comprises: mean values of the intervals of the heart beats, standard deviations of the intervals of the heart beats, skewness of the intervals of the heart beats, kurtosis of the intervals of the heart beats, NN50 metric values, pNN50 index, standard deviations of the differences between adjacent intervals of the heart beats, square root of the mean squared difference of the differences between adjacent intervals of the heart beats, alan factor, very low frequencies of the intervals of the heart beats, high frequencies of the intervals of the heart beats, and a ratio of low frequencies LF and high frequencies HF of the intervals of the heart beats.

6. The automatic sleep apnea detecting device of claim 1, wherein said sleep apnea determining module is further configured to determine whether sleep apnea occurs by:

the first heart rate variability features automatically extracted by the residual error neural network and the second heart rate variability features manually extracted are fused into a feature vector and input into a layer of fully-connected neural network classifier; the feature vector passes through the classifier to obtain an unnormalized discrimination value x, and the discrimination value is converted into a pseudo probability that the patient has sleep apnea in the current time period through an output function sigmoid (x); when the discrimination value x is not negative, the false probability is greater than or equal to 50%, and the discrimination is carried outPredicting the label value when the patient has sleep apnea in the current time period

Equal to 1; when the discrimination value x is negative, the pseudo probability is less than 50%, the patient is judged not to have sleep apnea in the current time period, and the label value is predicted

Equal to 0.

Images