TWM590434U - Detection device for obstructive sleep apnea - Google Patents

Abstract

A detection device for obstructive sleep apnea (OSA) is provided. The detection device includes a process, a storage medium, and a transceiver. The transceiver obtains a first ECG, wherein the first ECG includes a first data set corresponding to non-OSA and a second data set corresponding to the OSA. The processor accesses and executes a plurality of modules in the storage medium, wherein the plurality of modules include a training module and a detection module. The training module uses the first data set and the second data set as training data to train a machine learning model. The detection module obtains, via the transceiver, a second ECG of a subject, and determines, according to the machine learning model and the second ECG, whether the subject has the OSA.

Description

Device for detecting symptoms of obstructive sleep apnea

The disclosure relates to a detection device, and particularly to a detection device for obstructive sleep apnea (OSA) symptoms.

The OSA symptom is a common sleep disorder, which is a symptom of complete or partial upper airway obstruction caused by collapse of the pharynx during sleep, resulting in apnea or weakening. At present, the main basis for diagnosing obstructive sleep apnea is polysomnography (PSG). During the PSG examination, the subject must go to a sleep laboratory or sleep center for one night, under the supervision of the nursing staff, apply electrode patches on the head, corners of the eyes, chin, heart, and legs, and put them on the chest Put a sensor belt on the abdomen, put a blood oxygen meter on the finger, put a breath sensor on the nose and nose, and put a sphygmomanometer on the arm to record the sleep physiological data throughout the night. However, not all patients have time to spend the night in a sleep laboratory or sleep center, and patients with OSA symptoms do not have apnea events every night. Therefore, even if the patient goes to a sleep laboratory or sleep center overnight to measure sleep physiology data, the sleep physiology data measured that night may not be able to help the doctor diagnose the patient's OSA symptoms. According to this, how to propose a simplified and convenient method for diagnosing sleep disorders is one of the goals of those skilled in the art.

The present disclosure provides an OSA symptom detection device that can diagnose whether a subject suffers from OSA symptoms.

The disclosed device for detecting obstructive sleep apnea symptoms includes a processor, a storage medium, and a transceiver. The transceiver obtains a first electrocardiogram, where the first electrocardiogram includes a first data set corresponding to non-obstructive sleep apnea symptoms and a second data set corresponding to obstructive sleep apnea symptoms. The storage medium stores multiple modules. The processor is coupled to the storage medium and the transceiver, and accesses and executes multiple modules. The multiple modules include a training module and a detection module. The training module uses the first data set and the second data set as training data to train the machine learning model. The detection module obtains the second electrocardiogram of the subject through the transceiver, and determines whether the subject suffers from obstructive sleep apnea symptoms according to the machine learning model and the second electrocardiogram.

In an embodiment of the present disclosure, the above-mentioned first electrocardiogram and second electrocardiogram correspond to non-apnea events.

In an embodiment of the present disclosure, the aforementioned machine learning model is a convolutional neural network model.

In an embodiment of the present disclosure, the above-mentioned training module generates a corresponding cardiogenic respiratory signal according to the first electrocardiogram, measures a plurality of RR intervals according to the first electrocardiogram, and at least according to the plurality of RR intervals and at least the cardiogenic respiratory signal One of them generates second training data and trains the support vector machine model according to the second training data.

In an embodiment of the present disclosure, the above detection module determines whether the subject suffers from obstructive sleep apnea symptoms based on the support vector machine model, the machine learning model, and the second electrocardiogram.

In an embodiment of the present disclosure, the training module described above extracts multiple features from multiple RR intervals or cardiogenic respiratory signals, and generates second training data according to the multiple features.

In an embodiment of the present disclosure, the above-mentioned multiple features are associated with at least one of the following: the average value of the RR interval, the correlation coefficient of the second or third sequence of the RR interval, and the number of RR interval pairs, where the RR interval pair Including two adjacent RR intervals, and the time interval between the two RR intervals exceeds 50 milliseconds, the standard deviation of the two adjacent RR intervals, the normalized extreme low frequency range power of the RR interval, cardiogenic respiratory signal The normalized ultra-low frequency range power, the normalized low frequency range power of the cardiogenic respiratory signal, and the normalized high frequency range power of the cardiogenic respiratory signal.

In an embodiment of the present disclosure, the above-mentioned detection module issues a warning through the transceiver in response to determining that the subject suffers from obstructive sleep apnea symptoms.

Based on the above, the ECG of the subject under normal conditions (ie, no non-apnea event) can be used to diagnose whether the subject suffers from OSA symptoms.

In order to make the contents of the disclosure easier to understand, the following specific embodiments are taken as examples on which the disclosure can indeed be implemented. In addition, wherever possible, elements/components/steps with the same reference numerals in the drawings and embodiments represent the same or similar components.

FIG. 1 illustrates a schematic diagram of an OSA symptom detection device 100 according to an embodiment of the present disclosure. The detection device 100 includes a processor 110, a storage medium 120 and a transceiver 130.

The processor 110 is, for example, a central processing unit (CPU), or other programmable general-purpose or special-purpose micro-control unit (MCU), microprocessor (microprocessor), and digital signal processing (Digital signal processor, DSP), programmable controller, application specific integrated circuit (ASIC), graphics processing unit (GPU), arithmetic logic unit (ALU) , Complex programmable logic device (CPLD), field programmable gate array (FPGA) or other similar components or a combination of the above components. The processor 110 may be coupled to the storage medium 120 and the transceiver 130, and access and execute multiple modules and various application programs stored in the storage medium 120.

The storage medium 120 is, for example, any type of fixed or removable random access memory (RAM), read-only memory (ROM), flash memory (flash memory) , A hard disk drive (HDD), a solid state drive (SSD), or a similar component or a combination of the above components, and is used to store multiple modules or various applications that can be executed by the processor 110. In this embodiment, the storage medium 120 may store a plurality of modules including a training module 121 and a detection module 122, the functions of which will be described later.

The transceiver 130 transmits and receives signals in a wireless or wired manner. The transceiver 130 may also perform operations such as low noise amplification, impedance matching, frequency mixing, up or down frequency conversion, filtering, amplification, and the like. The transceiver 130 may be used to receive a first electrocardiogram as training data or a second electrocardiogram measured from a subject. For example, the transceiver 130 may be communicatively connected to an electrode attached to the subject, where the electrode may sense the subject's heartbeat to generate the subject's second electrocardiogram. For example, the transceiver 130 may use, for example, global system for mobile communication (GSM), personal handy-phone system (PHS), and code division multiple access (CDMA) System, wideband code division multiple access (WCDMA) system, long term evolution (LTE) system, global interoperability for microwave access (WiMAX) system, wireless fidelity (wireless Fidelity, Wi-Fi) system or Bluetooth (Bluetooth) and other communication technologies receive the first electrocardiogram as training data, or the second electrocardiogram measured by electrodes attached to the chest of the subject.

The first electrocardiogram as training data may include a first data set corresponding to non-OSA symptoms and a second data set corresponding to OSA symptoms, where the first data set is measured, for example, from a person who has not suffered from OSA symptoms in normal conditions (ie : The person’s electrocardiogram due to the heartbeat under the apnea event, and the second data set is, for example, to measure the heartbeat of the patient with self-severe OSA symptoms under normal conditions (ie: the patient has no apnea event) The resulting ECG. In this embodiment, the first data set may include multiple one-minute ECGs corresponding to non-OSA symptoms, and no apnea event occurs during each ECG measurement period. The second data set may include multiple one-minute ECGs corresponding to OSA symptoms, and no apnea event occurs during each ECG measurement period.

The training module 121 can use the first data set and the second data set as training data to train the machine learning model, wherein the trained machine learning model can be used to determine whether the subject suffers from OSA according to the second electrocardiogram measured from the subject symptom. It is worth noting that the training data in this embodiment may be a time-domain signal that has not been transformed by wavelet, instead of a time-frequency signal that has been transformed by wavelet. Since the dimension of the time-domain signal is lower than that of the time-frequency signal, the training module 121 that uses the time-domain signal instead of the time-frequency signal to train the machine learning model will cost less computing power.

In one embodiment, the first electrocardiogram as training data further includes a third data set representing an electrocardiogram interfered by noise. The training module 121 may use the first data set, the second data set, and the third data set as training data to train the machine learning model. The machine learning model trained from the first data set, the second data set, and the third data set can not only determine whether the subject suffers from OSA symptoms based on the second ECG measured from the subject, but also based on the second ECG The second ECG data received by the transceiver 130 may be interfered by noise. Therefore, when determining the severity of the OSA symptom of the subject, the arithmetic module 122 may first filter out the data interfered by noise.

After training the machine learning model, the detection module 122 can determine whether the subject suffers from OSA symptoms according to the machine learning model and the second electrocardiogram measured from the subject.

In one embodiment, if the detection module 122 determines that the subject suffers from OSA symptoms, the detection module 122 may issue a warning through the transceiver 130 to alert the subject that he has suffered from OSA symptoms. For example, the detection module 122 may send a warning message to the subject's mobile device through the transceiver 130 after determining that the subject has OSA symptoms.

The aforementioned machine learning model is, for example, a convolutional neural network (CNN) model. The biggest difference between the convolutional neural network and the traditional multi-layer perception network is that the convolutional neural network has more convolutional layers and pooling layers. These two layers give the convolutional neural network the ability to extract the characteristics of the input signal. The design of the convolutional layer has many characteristics. The first feature is local perception. In the traditional neural network, each neuron must be connected with each sampling point, so a large amount of weight is required, which makes the difficulty of training the network extremely high. In the convolutional neural network, the weight of each neuron is the same as the size of the convolution kernel, so it is equivalent to that each neuron is only connected to the corresponding part of the sampling point, which can greatly reduce the number of weights. . A relatively small number of weights can reduce the risk of overfitting. The second feature is the weight sharing mechanism. Convolutional neural network is to train and update the optimal weight of the convolution kernel through the back propagation error algorithm, but in the process of convolution, the weight of the convolution kernel will not change. The third feature is the multi-convolution kernel. If only one convolution kernel is used, only some features of the signal can be extracted. If multiple convolution kernels are used, multiple features of the input signal can be extracted. The more the number of convolutional layers, the more features the convolutional neural network can extract.

After the input signal undergoes a non-linear transformation by the convolution layer and the activation function, a feature map can be generated. The most important function of the activation function is to introduce the nonlinearity of the neural network, because if no activation function is added, the convolutional layer and the fully connected layer are only simple linear operations, and there is still no solution to the problem of linear inseparability.

In order to reduce the dimension of the features extracted by the convolution operation and increase the speed of the learning process, the convolution layer will be followed by a pooling layer. The pooling layer is a method of compressing feature maps and retaining important information. The sampling method adopted by the pooling layer may include a max pooling method or a mean pooling method. The maximum pooling method is to select the maximum value in the pooling window as the sampling value. The average pooling method is to add all the values in the pooling window and take the average as the sampling value. The feature map after pooling still retains the maximum possibility of local range comparison. In other words, the pooled information can focus on whether there are matching features in the feature map, rather than on the location of these features. Therefore, compared with the traditional neural network, the convolutional neural network can determine whether a feature is included in the feature map, without considering the location of the feature. Therefore, even if the feature of the input signal is shifted, the convolutional neural network can recognize the feature. After the pooling layer, the fully connected layer integrates highly abstracted features after multiple convolutions and pooling. Then, the output layer outputs a corresponding probability for each classification, in which the total probability of all classifications is 1.

2 illustrates a schematic diagram of a convolutional neural network model 200 according to an embodiment of the present disclosure, where the convolutional neural network model 200 is a form of a machine learning model trained by the training module 121. The convolutional neural network model 200 may include an input layer 220, a convolutional layer 231, a pooling layer 232, a convolutional layer 241, a pooling layer 242, a fully connected layer 251, a fully connected layer 252, and an output layer 260. The input data 210 of the convolutional neural network model 200 shown in FIG. 2 is, for example, a second electrocardiogram measured from a subject, and the output data 270 of the convolutional neural network model 200 represents the judgment result of whether or not suffering from OSA symptoms.

的卷積核。經過卷積層231的輸入資料210會轉變為128個尺寸為

的特徵圖。接著，池化層232使用尺寸為

的滑動視窗對卷積層231輸出的特徵圖進行取樣以產生128個尺寸為

的特徵圖。卷積層241包括64個尺寸16

的卷積核。卷積層241進一步地對池化層232輸出的特徵圖進行卷積運算以產生64個尺寸為

的特徵圖。接著，池化層242使用尺寸為

的滑動視窗對卷積層241輸出的特徵圖進行取樣以產生64個尺寸為

的特徵圖。而後，池化層242輸出的特徵圖依序地輸入至具有128個神經元的全連接層251以及具有64個神經元的全連接層252。輸出層260可根據Softmax激活函數來計算全連接層252的輸出的對應於罹患OSA症狀的第一機率以及對應於未罹患OSA症狀的第二機率。若第一機率大於第二機率，則偵測模組122可判斷輸入資料210對應於罹患OSA症狀的患者。反之，若第一機率小於或等於第二機率，則偵測模組122可判斷輸入資料210對應未罹患OSA症狀的人員。 In this embodiment, the input data 210 is a second electrocardiogram of 1 minute, and the input data 210 includes 6,000 sampling points sampled at a sampling frequency of 100 Hz. Convolution layer 231 includes 128 dimensions of

Convolution kernel. The input data 210 through the convolution layer 231 will be transformed into 128 sizes of

Feature map. Next, the pooling layer 232 uses a size of

Of the sliding window samples the feature map output by the convolution layer 231 to produce 128 dimensions of

Feature map. Convolutional layer 241 includes 64 sizes 16

Convolution kernel. The convolution layer 241 further performs a convolution operation on the feature map output by the pooling layer 232 to generate 64 sizes of

Feature map. Next, the pooling layer 242 uses a size of

The sliding window samples the feature map output by the convolution layer 241 to produce 64 dimensions of

Feature map. Then, the feature map output by the pooling layer 242 is sequentially input to the fully connected layer 251 with 128 neurons and the fully connected layer 252 with 64 neurons. The output layer 260 may calculate the first probability corresponding to the OSA symptom and the second probability corresponding to the OSA symptom output of the fully connected layer 252 according to the Softmax activation function. If the first probability is greater than the second probability, the detection module 122 may determine that the input data 210 corresponds to a patient suffering from OSA symptoms. Conversely, if the first probability is less than or equal to the second probability, the detection module 122 may determine that the input data 210 corresponds to a person who does not suffer from OSA symptoms.

It is worth noting that the activation function used by the output layer 260 may be, for example, a softmax function, a sigmoid function, a hyperbolic tangent function, or a rectified linear unit (ReLU) function, and the disclosure is not limited thereto.

In one embodiment, the training module 121 can further generate a support vector machine (SVM) model. The detection module 122 can determine whether the subject suffers from OSA symptoms according to the support vector machine model, the machine learning model, and the second electrocardiogram measured from the subject. For example, if at least one of the support vector machine model and the machine learning model determines that the subject suffers from OSA symptoms, the detection module 122 may respond to at least one of the support vector machine model and the machine learning model to determine The test subject suffers from OSA symptoms and outputs a judgment result representing that the subject suffers from OSA symptoms.

The training module 121 can train the aforementioned support vector machine model according to the first electrocardiogram. The training module 121 may generate a corresponding ECG-derived respiration (EDR) signal according to the first electrocardiogram as training data. On the other hand, the training module 121 may measure multiple RR intervals according to the first electrocardiogram. Then, the training module 121 may generate second training data for training the support vector machine model according to at least one of the cardiogenic respiratory signal and the plurality of RR intervals, and then train the aforementioned training data according to the second training data Support vector machine model. Specifically, the training module 121 may extract multiple features from the cardiogenic respiration signal or multiple RR intervals, and generate second training data according to the features. The plurality of features are associated with, for example, the average value of the RR interval, the correlation coefficient of the second or third sequence of the RR interval, the number of RR interval pairs (the RR interval pair includes two adjacent RR intervals, and the two RR intervals Time interval of more than 50 milliseconds), the standard deviation of two adjacent RR intervals, the normalized very low frequency power (VLFP) of the RR interval, and the normalized very low frequency range of the cardiogenic respiratory signal Power, normalized low frequency power (LFP) of cardiogenic respiratory signals or normalized high frequency power (HFP) of cardiogenic respiratory signals, but the disclosure is not limited thereto. VLFP is approximately between 0.003-0.04 Hz. The above LFP is approximately between 0.04-0.15 Hz and HFP is approximately between 0.15-0.4 Hz.

FIG. 3 illustrates a flowchart of a method for detecting OSA symptoms according to an embodiment of the present disclosure, wherein the detection method is implemented by the detection device 100 shown in FIG. 1, for example. In step S301, a first electrocardiogram is obtained, where the first electrocardiogram includes a first data set corresponding to non-obstructive sleep apnea symptoms and a second data set corresponding to obstructive sleep apnea symptoms. In step S302, the first data set and the second data set are used as training data to train a machine learning model. In step S303, the second electrocardiogram of the subject is obtained. In step S304, it is determined whether the subject suffers from obstructive sleep apnea symptoms according to the machine learning model and the second electrocardiogram.

In summary, the present disclosure can use electrocardiograms of patients with non-OSA symptoms and patients with OSA symptoms under normal conditions (ie, no apnea event) as training materials to train machine learning models. The trained machine learning model can determine whether the subject suffers from OSA symptoms when the subject is under normal conditions. In other words, even if the subject suffering from OSA symptoms did not have an apnea event, the disclosure can still accurately predict the subject suffering from OSA symptoms based on the subject’s electrocardiogram and further prompt the subject to go to Hospital examination. In this way, the subject's OSA symptoms can be detected and treated early.

100‧‧‧detection device 110‧‧‧ processor 120‧‧‧storage media 121‧‧‧Training Module 122‧‧‧detection module 130‧‧‧Transceiver 210‧‧‧Enter data 220‧‧‧ input layer 231, 241‧‧‧ Convolutional layer 232, 242‧‧‧ Pooling layer 251, 252‧‧‧ fully connected layer 260‧‧‧Output layer 270‧‧‧ Output data S301, S302, S303, S304 ‧‧‧ steps

FIG. 1 is a schematic diagram of an OSA symptom detection device according to an embodiment of the present disclosure. FIG. 2 illustrates a schematic diagram of a convolutional neural network model according to an embodiment of the present disclosure. FIG. 3 illustrates a flowchart of a method for detecting OSA symptoms according to an embodiment of the present disclosure.

100‧‧‧detection device

110‧‧‧ processor

120‧‧‧storage media

121‧‧‧Training Module

122‧‧‧detection module

130‧‧‧Transceiver

Claims (8)

A device for detecting symptoms of obstructive sleep apnea, including: The transceiver obtains a first electrocardiogram, wherein the first electrocardiogram includes a first data set corresponding to non-obstructive sleep apnea symptoms and a second data set corresponding to the obstructive sleep apnea symptoms; Storage media to store multiple modules; and A processor is coupled to the storage medium and the transceiver, and accesses and executes the plurality of modules, wherein the plurality of modules includes: A training module, using the first data set and the second data set as training data to train a machine learning model; and The detection module obtains a second electrocardiogram of the subject through the transceiver, and determines whether the subject suffers from the obstructive sleep apnea symptom according to the machine learning model and the second electrocardiogram. The detection device according to item 1 of the patent application scope, wherein the first electrocardiogram and the second electrocardiogram correspond to non-apnea events. The detection device according to item 1 of the patent application scope, wherein the machine learning model is a convolutional neural network model. The detection device according to item 1 of the patent application scope, wherein the training module generates a corresponding cardiogenic respiratory signal according to the first electrocardiogram, measures a plurality of RR intervals according to the first electrocardiogram, and according to the At least one of the plurality of RR intervals and the cardiogenic respiratory signal generates second training data, and trains a support vector machine model according to the second training data. The detection device according to item 4 of the patent application scope, wherein the detection module determines whether the subject suffers from the support vector machine model, the machine learning model, and the second electrocardiogram Obstructive sleep apnea symptoms. The detection device according to item 4 of the patent application scope, wherein the training module extracts a plurality of features from the plurality of RR intervals or the cardiogenic respiratory signal, and generates the Second training data. The detection device according to item 6 of the patent application scope, wherein the plurality of features are associated with at least one of the following: The average value of the RR interval, the correlation coefficient of the second or third sequence of the RR interval, and the number of RR interval pairs, wherein the RR interval pair includes two adjacent RR intervals, and the time interval between the two RR intervals More than 50 ms, standard deviation of two adjacent RR intervals, normalized low frequency range power of RR interval, normalized low frequency range power of cardiogenic respiratory signal, normalized low frequency of cardiogenic respiratory signal Range power and normalized high frequency range power of cardiogenic respiratory signals. The detection device according to item 1 of the patent application scope, wherein the detection module issues a warning through the transceiver in response to determining that the subject suffers from the obstructive sleep apnea symptom.

Images