KR20230085526A - Apparatus and method for classifying breathing state during sleep using biosignals - Google Patents

Abstract

The present invention relates to an apparatus and method for classifying sleep breathing conditions using bio-signals, and more particularly, to a bio-signal measuring unit for measuring an ECG signal during sleep of a user, a pre-processing unit for pre-processing the ECG signal, and the pre-processed ECG It includes a feature extraction unit that extracts characteristics of the signal and a diagnosis unit that diagnoses the user's breathing state using the feature, and the feature extraction unit and the diagnosis unit are composed of deep learning and can diagnose OSA only with ECG signals. .

Description

Apparatus and method for classifying breathing state during sleep using biosignals}

The present invention relates to an apparatus and method for classifying sleep breathing conditions using biosignals, and more specifically, to an apparatus and method for measuring the severity of sleep apnea of a user by measuring an ECG signal during the user's sleeping time and providing the result to the user. .

Obstructive sleep apnea (OSA) is one of the common sleep disorders suffered by both men and women due to the increase in aging and obesity due to prolonged sleep as industrialization and social development progresses. It causes complications related to the cardiovascular system such as coronary artery disease, arrhythmia, and pulmonary hypertension. In addition, lack of sleep causes drowsiness during the daytime, resulting in social problems such as drowsy driving and reduced productivity.

In general, obstructive sleep apnea is diagnosed through polysomnography, which can determine the overall state of sleep and sleep disorders. Polysomnography is performed in an examination room with various sensors attached to the user's body, and because experts observe an average of 7 to 8 hours of sleep, there are restrictions on location and money. Such limitations make polysomnography difficult for all potentially obstructive sleep apnea patients.

In addition, for obstructive sleep apnea testing, it is essential to measure sleep time and detect apnea events. Therefore, a long time can be misjudged as sleep time, and the Apnea-Hypopnea Index (AHI) can be calculated smaller than it actually is.

In order to solve the above problem, various technologies have been proposed in the past, but it is necessary to measure various bio-signals with a plurality of sensors, and based on average data to be suitable for a large number of users rather than a customized evaluation method for the user. However, there is still a problem in that these methods do not accurately provide the apnea-hypopnea index.

Republic of Korea Patent Publication No. 10-2020-0139049 (published on December 11, 2020)

The present invention is to solve the problem of the above technical problem, and an object of the present invention is to provide a method for measuring the apnea-hypopnea index to the user using only a single signal of an electrocardiogram (ECG).

In another aspect, the present invention measures an accurate apnea-hypopnea index by providing a customized obstructive sleep apnea evaluation model for each user using deep learning, and estimates and provides the severity of obstructive apnea based on the apnea-hypopnea index. there is

An apparatus for classifying sleep breathing conditions according to an embodiment of the present invention for solving the above technical problem is a bio-signal measuring unit for measuring an ECG signal during sleep of a user, a pre-processing unit for pre-processing the ECG signal, and the pre-processed ECG signal. It includes a feature extraction unit for extracting features of and a diagnosis unit for diagnosing the user's respiratory state using the features, and the feature extraction unit and the diagnosis unit may be configured with deep learning.

In addition, the pre-processing unit includes a segmentation unit that subdivides the ECG signal into predetermined units using a discrete sliding window method, a labeling unit that labels each region of the ECG signal segmented in the segmentation unit, and the labeling unit. and a filtering unit for filtering the classified labels, and the labeling unit may perform labeling by grouping regions before and after the region to be diagnosed by the diagnosis unit among the regions subdivided by the subdivision unit.

In addition, the respiration state may be re-diagnosed by resetting a subdivision unit of the subdivision unit to a unit smaller than the preset unit according to the reliability of the diagnosis unit.

The feature extraction unit may include a peak detector for detecting a peak of the R wave in the ECG signal, a feature generator for generating a feature based on the peak of the R wave, and a resizing unit for resizing the feature.

In addition, the characteristics are the interval from the R wave to the next R wave in the ECG (RRI), ECG-Derived Respiration (EDR), and the amplitude of the ECG (Amplitude of ECG, AMP), and the resizing unit interpolates The RRI may perform oversampling and the EDR may perform downsampling.

In addition, the diagnosis unit includes an OSA detection unit for detecting obstructive sleep apnea (OSA) using the feature, a sleep time estimation unit for estimating pure sleep time using the feature, and an apnea-lowering using the OSA and the sleep time. It may include an AHI calculation unit that calculates a respiratory index (AHI) and a respiratory state estimation unit that estimates the severity of sleep apnea based on the AHI.

In addition, the deep learning is composed of a combination of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), and the OSA and AHI can calculate the probability of each breath using softmax.

In addition, the convolutional neural network may be learned using the measurement result of the diagnosis unit and the ECG signal.

A sleep breathing state classification method according to another embodiment of the present invention for solving the above technical problem includes measuring an ECG signal in a sleep state, dividing the ECG signal into predetermined units, and dividing the ECG signal Extracting a feature of and diagnosing a respiratory state using the feature, and extracting the feature and diagnosing the respiratory state may use a classification model composed of deep learning.

In addition, the dividing into predetermined units may include dividing the ECG signal into predetermined units using a discrete sliding window method, labeling each of the divided regions, and among the labeled bells. It may include grouping labels before and after the label to be measured and setting them as one group, and filtering the group.

In addition, the respiratory state may be re-diagnosed by further subdividing the subdivided unit into smaller units than a predetermined unit according to the reliability of the respiratory state.

In addition, the step of extracting the feature may include detecting a peak of the R wave in the divided ECG signal, extracting a feature based on the peak of the R wave, and resizing the feature.

In addition, the characteristics are the interval from R wave to next R wave in the electrocardiogram (RRI), ECG-Derived Respiration (EDR), and amplitude of ECG (amplitude of ECG, AMP), the resizing step Using an interpolation method, the RRI may perform oversampling and the EDR may perform downsampling.

In addition, diagnosing the respiratory state includes detecting obstructive sleep apnea (OSA) and pure sleep time using the characteristics, calculating an apnea-hypopnea index (AHI) based on the OSA and the sleep time. and estimating the severity of sleep apnea.

In addition, the deep learning is composed of a combination of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), and the OSA and AHI can calculate the probability of each breath using softmax.

In addition, diagnosing the respiratory state may further include performing learning of the convolutional neural network using a measurement result and the ECG signal.

According to the apparatus and method for classifying sleep breathing conditions according to an embodiment of the present invention described above, obstructive apnea can be diagnosed using only a single ECG signal, and the size of the apparatus can be greatly reduced.

In addition, by diagnosing obstructive apnea using deep learning, it is possible to provide a device and method for measuring obstructive apnea customized for each user.

In addition, the severity of obstructive apnea can be more accurately calculated by calculating the obstructive apnea-hypopnea index by separating the time of occurrence of obstructive apnea and actual sleep time.

In addition, a more accurate sleep time can be calculated by changing the measurement unit time of obstructive apnea.

1 is a block diagram of an apparatus for classifying a sleep breathing state according to an embodiment of the present invention.
2 is a diagram illustrating a method of dividing an ECG signal in a sleep breathing state classification apparatus according to an embodiment of the present invention.
3 is a diagram of a structure of deep learning used in a sleep breathing state classification apparatus according to an embodiment of the present invention.
4 is a flowchart of a sleep breathing state classification method according to an embodiment of the present invention.
5 is a flow chart specifying the preprocessing step of the sleep breathing state classification method according to an embodiment of the present invention.
6 is a flow chart specifying a feature extraction step of a sleep breathing state classification method according to an embodiment of the present invention.
7 is a flow chart specifying steps for specifying a method for diagnosing a breathing state of a sleep breathing state classification method according to an embodiment of the present invention.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail.

However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention.

Terms used in this application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this application, the terms "include" or "have" are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other features It should be understood that the presence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and unless explicitly defined in the present application, they should not be interpreted in an ideal or excessively formal meaning. don't

Hereinafter, with reference to the accompanying drawings, preferred embodiments of the present invention will be described in more detail. In order to facilitate overall understanding in the description of the present invention, the same reference numerals are used for the same components in the drawings, and redundant descriptions of the same components are omitted.

1 is a block diagram of an apparatus for classifying a sleep breathing state according to an embodiment of the present invention.

An apparatus for classifying sleep breathing conditions according to an embodiment of the present invention may include a bio-signal measuring unit 100, a pre-processing unit 200, a feature extraction unit 300, and a diagnosis unit 400.

The bio-signal measurer 100 may directly contact any part of the user's body to measure the user's bio-signal in a non-invasive manner, and in particular, may be a sensor that measures an electrocardiogram (ECG) signal.

In general, ECG signals can be obtained by attaching a sensor to any part of the body that can measure ECG using electrodes, such as wrists, chests, and ankles, without limitation. An ECG signal can be acquired without difficulty even during sleep by using an electronic device.

An embodiment of the present invention discloses a method for estimating obstructive sleep apnea (OSA) using only an ECG signal, but is not limited thereto, and if necessary, a pulse wave (PPG), Accuracy can be increased by adding various information such as oxygen saturation (SpO ₂ ), user movement information using an acceleration sensor, and heart rate variability (HRV). It is natural that you can

The pre-processing unit 200 performs data pre-processing to estimate the OSA using the ECG signal, and may include a segmentation unit 210, a labeling unit 220, and a filtering unit 230.

The segmentation unit 210 may segment the ECG signal measured by the bio-signal measuring unit 100 into predetermined units using a discrete sliding window method.

In one embodiment of the present invention, the segmentation unit 210 may generate a batch by default in 1 minute units, but may generate batches in various size units such as 10 seconds, 30 seconds, and 5 minutes as needed. there is.

For example, if the user evaluates that the OSA is high in the previous measurement results or the reliability of the measured OSA is low, the batch unit may be adjusted from 1 minute to 30 seconds or 10 seconds, and vice versa. In the case of low measured or highly reliable users, the batch unit can be increased from 1 minute to 5 minutes and 10 minutes.

In addition, in order to improve measurement reliability, the accurate OSA duration, duration, and measurement accuracy can be improved by re-measuring the batch in which OSA is detected as a result of ECG signal analysis by dividing it into 30 seconds, 10 seconds, etc. can

That is, the representative embodiment of the present invention discloses a method of generating a batch by subdividing the segmentation unit 210 by 1 minute, but is not limited thereto, and the reliability of the measured situation, previous measurement data, and measured ECG signal for each user. It is possible to dynamically change the generation unit of a batch according to various situations.

The labeling unit 220 may perform labeling as an epoch for applying the arrangement generated by the segmentation unit 210 to the diagnosis unit 400 .

When OSA is detected by directly applying the arrangement generated by the subdivision unit 210 to the diagnosis unit 400, there is a problem in that the accuracy of OSA detection may be reduced because a non-continuous signal is applied.

FIG. 2 is an example of an operation method of the labeling unit 210. Referring to FIG. 2, the operation of the labeling unit 210 will be illustrated.

For OSA detection, more accurate measurement is possible only when the measurement is performed in a continuous time series. To this end, the subdivision unit 210 may utilize a method of performing labeling by tying together batches before and after OSA measurement.

In one embodiment of the present invention, in order to secure the continuity of a time-series ECG signal, it is necessary to basically consider the signals that follow the selected signal, and for this purpose, in one embodiment of the present invention, the signal For the ECG signal divided into minutes, a total of 5 minutes of time, 2 minutes before and 2 minutes after the batch selected for OSA measurement, can be set as one label.

For example, for ECG signals measured for 10 minutes, 10 batches of 1 minute each may be created and numbers 1 to 10 may be sequentially assigned from the first batch. For example, if signal No. 3 is measured thereafter, the first two signals 1 and 2, and the two after No. 3, Nos. 4 and 5, are grouped as one group, and ECG signals in 5-minute units from No. 1 to No. 5 are labeled as one label. can be set to

Similarly, when signal 4 is measured, signals 2 and 3 in front and signals 5 and 6 in the back can be grouped together and patches 2, 3, 4, 5, and 6 can be labeled as one epoch.

The filtering unit 230 may perform filtering in units of labels labeled in the labeling unit 220 .

In one embodiment of the present invention, the filtering unit 230 performs an operation using a Finite Impulse Response Filter (FIR) filter, which is a type of low pass filter, but is not limited thereto, and a Butterworth filter ), a filter such as a powerline filter may be used.

In one embodiment of the present invention, the filtering unit 230 may perform filtering using data provided from PhysioNet.

The feature extraction unit 300 may extract features from the ECG signal data preprocessed by the preprocessor 200, and may include a peak detection unit 310, a feature generator 320, and a resizing unit 330. .

The peak detection unit 310 may detect the peak of the R wave with respect to the ECG signal within the label filtered by the filtering unit 230.

The electrical activity phase of the heart can be largely divided into atrial depolarization, ventricular depolarization, and ventricular repolarization phases, and each of these phases can be reflected in the form of several waves called P waves, QRS waves, and T waves. In the example, the R wave among the types of P wave, QRS wave, and T wave is detected and used.

In one embodiment of the present invention, the peak detection unit 310 detects the R wave using Neurokit, but is not limited thereto and is not limited to use as long as it can detect the R wave.

The feature generator 320 may extract features of the ECG signal.

The feature generation unit 320 performs feature extraction in units of five batches labeled in the labeling unit 220, and may extract features in units of five batches, which is 5 minutes, as in the representative embodiment of the present invention.

The feature generation unit 320 has three features of RRI, which is the interval between R waves extracted by the peak detection unit 310, ECG-Derived Respiration (EDR), and Amplitude of ECG (AMP) can be extracted.

As described above, the present invention can detect OSA only with three characteristic signals of RRI, EDR, and AMP that can be extracted from an ECG signal, and can also use the characteristics of additional signals selectively measured as needed. am.

One batch consists of a plurality of samples, and the plurality of samples may consist of samples of the three characteristics (RRI, EDR, AMP). Resizing may be performed using these samples to perform more accurate OSA detection.

The resizing unit 330 may change the size of the feature extracted from the feature generator 330 so that the diagnosis unit 400 can perform a more accurate diagnosis.

In an embodiment of the present invention, the resizing unit 330 may resize the features generated by the feature generator 320 using an interpolation method, perform oversampling based on the interpolation method for RRI, and interpolation method for EDR. Downsampling can be performed based on .

The resizing unit 330 is not an essential element of the sleep breathing state classification device of the present invention, but is implemented for more accurate detection of OSA, and if the detection performance can be improved, the method is not limited and the feature is resized. can do.

The diagnosis unit 400 may calculate OSA and Apnea-Hypopnea Index (AHI) based on the features extracted by the feature extraction unit 300, and may calculate OSA severity based on the AHI.

The diagnosis unit 400 may include an OSA detection unit 410, a sleep time estimation unit 420, an AHI calculation unit 430, and a breathing state estimation unit 440.

The OSA detection unit 410 may perform OSA detection based on the features extracted by the feature extraction unit 300 .

Referring back to FIG. 2 , among batches 1 to 10, batches 3, 6, and 8 illustrate an apnea state, not normal breathing. At this time, when the inspection is performed on batch No. 3, as described in the labeling unit 220, the feature extraction unit 300 adds the features in the feature extraction unit 300 based on a total of five batches (Batches 1 to 5) by adding two batches before and after. is extracted, and when the extracted feature is applied to the OSA detection unit 410, the OSA detection unit 410 outputs a result of normal when breathing is normal and abnormal when apnea is detected, and the label centered on batch 3. The result can be output as abnormal, that is, in the OSA state.

In the case of performing the test on batch No. 4, batches No. 2 to 6 are set as one label by adding two batches before and after, and since batch No. 4 is normal, a normal result can be output.

The sleep time estimator 420 may estimate pure sleep time excluding wake time, such as temporarily waking up, by the OSA detected by the OSA detector 410 .

The sleep time estimator 420 may generally perform an operation by receiving a signal from the feature extraction unit 300 simultaneously with the OSA detection unit 410, and the time excluding the detection result of the OSA detection unit 410 is regarded as pure sleep time. can be estimated

In addition, when the OSA temporarily enters the wake state from sleep, it does not immediately enter the sleep state again, so it can be estimated as the pure sleep time by excluding the wake time additionally based on the experimental results.

In other words, the features extracted by the feature extraction unit 300 are applied to the OSA detection unit 410 and the sleep time estimation unit 420 to be classified into OSA and normal sleep, and in one embodiment of the present invention, a separate configuration , but may exist as a single component.

The AHI calculation unit 430 may calculate an Apnea-Hypopnea Index (AHI) based on the OSA and pure sleep time measured by the OSA detection unit 410 and the sleep time estimation unit 420.

Apnea generally means stopping breathing for more than 10 seconds, and hypopnea is a decrease in ventilation (respiratory volume) by more than 50% for more than 10 seconds. means average number of times.

Since batches are generated in units of 1 minute in the segmentation unit 210 according to an embodiment of the present invention, a total of 60 batches are generated in 1 hour, and 60 epochs can be generated in the same way in the labeling unit. The feature data of these 60 epochs is applied to the OSA detection unit 410 and the sleep time estimation unit 420 to measure the OSA and the pure sleep time to calculate the AHI, thereby allowing a more accurate AHI to be calculated.

In particular, in one embodiment of the present invention, AHI is calculated using pure sleep time, and there is a clear difference from a configuration in which AHI is calculated only with OSA as in the prior art, and such pure sleep time is included in the calculation of AHI. By doing so, a more accurate AHI can be calculated.

The respiratory state estimation unit 440 may classify the severity of OSA by receiving the AHI value of the AHI calculation unit 430 .

In general, an AHI of less than 5 is normal, 5 to 15 is mild, 15 to 30 is moderate, and more than 30 is classified as severe, and severity can be provided to users according to this classification.

In addition, although not disclosed in the drawings, the measurement results of the device for classifying sleep breathing conditions may be transmitted and stored in a storage device (not shown) such as a cloud or server using a wired or wireless communication network, and may be stored by a user in a medical institution such as a hospital or clinic. The stored information can be provided and used for medical purposes such as diagnosis and treatment.

3 is a diagram of a structure of deep learning used in a sleep breathing state classification apparatus according to an embodiment of the present invention.

In one embodiment of the present invention, OSA can be detected using deep learning, and deep learning can be configured by a combination of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN).

Time-series data classification of ECG data is performed using LSTM, and CNN can be used to determine OSA and pure sleep time. In addition, it goes without saying that features can be extracted using CNN.

In addition, for convenience, as shown in FIG. 1, the diagnosis unit 400 is composed of an OSA detection unit 410, a sleep time estimation unit 420, an AHI calculation unit 430, and a breathing state estimation unit 440 so as to be more easily understood according to convenience. ), but in reality, all of them can be processed and provided in one deep learning. It is also possible to perform AHI calculation and respiratory state estimation.

3 is an example in which only the configurations of the OSA detection unit 410 and the sleep time estimation unit 420 are implemented by deep learning.

If the number of layers increases, there is a possibility of congestion due to loss of gradient and excessive operation of the neural network.

In an embodiment of the present invention, batch-normalization may be used whenever passing through a 1-D convolution filter to minimize the possibility of loss of gradient and convergence problems.

In addition, dropout can be set to 0.7 to prevent overfitting according to data in deep learning.

In addition, it is possible to classify normal respiration and abnormal respiration (OSA) based on ECG characteristics (RRI, EDR, AMP) using softmax.

In one embodiment of the present invention, a method for performing respiratory classification based on deep learning using LSTM and CNN is exemplified, but is not limited thereto, and if respiratory classification can be performed, all RNNs, ANNs, GANs, machine learning, etc. It can be configured using an artificial neural network.

Deep learning in one embodiment of the present invention may be a model learned in advance using data provided from Physionet, data obtained through conventional polysomnography, etc., and measured in various forms of Hertz (Hz) in advance Respiration classification can be performed regardless of the hertz used for ECG measurement by performing learning with a training set.

In addition, since deep learning in one embodiment of the present invention has been completed in advance, the user can immediately use it for measurement, and the measurement result can be stored in a storage device (not shown) and used for additional learning of deep learning. there is.

That is, the deep learning of the present invention can form a customized model for each user as the use is repeated, and the accuracy of OSA measurement can be improved as the customized learning for each user is performed.

In addition, since multiple users measure using one device, there may be problems with deep learning learning. can provide

4 is a flowchart of a sleep breathing state classification method according to an embodiment of the present invention.

When the user wears the device for classifying sleep breathing conditions and performs sleep, the bio-signal measuring unit 100 may measure the ECG signal (S100).

The measured ECG signal may be divided into predetermined units (S200).

The features of the divided ECG signal are extracted (S300), and the respiratory state can be diagnosed using the extracted features (S400).

5 is a diagram for explaining step S200 in detail.

When the ECG signal is applied, the pre-processing unit 200 may perform pre-processing of dividing into predetermined units. To perform preprocessing, the measured ECG signal may be divided into preset units in the segmentation unit 210 (S210).

In principle, the divided unit is divided into 1 minute, but is not limited thereto, and if more specific evaluation is required according to the previous measurement result, it is subdivided into 30 seconds, 15 seconds, etc., or in the case of users with less sleep apnea By widening the interval to 5 minutes, 10 minutes, etc., the units that are dynamically divided can be adjusted to secure the ease of measurement and improve the degree of accuracy.

When the ECG signal is divided, labeling may be performed on each divided region (S220).

Since the ECG signal is a signal with temporal continuity, OSA detection also includes continuous time series, so the label used for the test must include continuity for more accurate measurement.

In one embodiment of the present invention, a region combining two regions before and after the divided signal to be measured may be labeled with one label so as to ensure continuity of the divided ECG signal.

For example, referring to FIG. 2, when the measurement period is divided into 10 areas, in order to inspect area 3, two areas in front of area 3, areas 1 and 2, and two areas after area 4, 5 Areas 1 to 5, which are burn areas, may be labeled with one label.

In order to inspect area 5 in the same way, the front two areas, areas 3 and 4, and the back two areas, areas 6 and 7, areas 3 to 7, may be labeled with one label.

When labeling is performed in the above manner, signal continuity can be ensured because signals 2, 3, 4, and 5 are equally included in each label when signal 4 is inspected after signal 3 is inspected.

When labeling of the ECG signal is completed, noise and the like may be filtered to improve measurement accuracy for each label (S230).

6 is a diagram for explaining the step S300 in detail.

The divided signal may be applied to the feature extractor 300 to extract features.

Each label is applied to the peak detection unit 310 to detect the R-peak of the ECG signal within the label (S310).

The ECG signal includes a plurality of waveforms such as P wave, Q wave, R wave, S wave, and T wave, and in one embodiment of the present invention, the R wave corresponding to the peak can be detected and used.

R-Peak (R wave) can be detected using Neurokit, and in addition to this, the operation can be performed using various methods that can analyze ECG signals capable of detecting R waves.

A plurality of features may be extracted using the detected R-peak (S320).

In one embodiment of the present invention, the plurality of features include R-peak R-peak Interver (RRI), which is an interval between R-peaks, ECG-Derived Respiration (EDR), and Amplitude of ECG, AMP) is used to detect OSA, but in addition to this, additional signals and features can be extracted and used together according to the needs of increased accuracy and environment.

The plurality of extracted features may be resized so that the diagnostic unit 400 can perform a more accurate diagnosis (S330).

For example, for the resizing, the resizing unit 330 may perform oversampling based on an interpolation method for RRI and downsampling based on an interpolation method for EDR.

7 is a diagram for explaining the step S400 in detail.

The OSA detection unit 410 may detect the presence or absence of OSA in the plurality of features for which resizing is completed (S410).

The OSA detecting unit 410 analyzes the characteristics of the ECG signal of the label located in the middle of the input label to classify it into normal respiration and abnormal respiration. can be detected.

In addition, for the plurality of features for which resizing is completed, the sleep time estimation unit 420 may estimate pure sleep time (S420).

Steps S410 and S420 may be performed simultaneously as steps of receiving and operating the features extracted by the feature extraction unit 300, or may be performed sequentially in the order of measuring the sleep time after OSA detection.

In addition, the steps S410 and S420 may be performed using deep learning composed of LSTM and CNN, and the operation S300 may also be performed using deep learning, if necessary.

AHI may be calculated based on the OSA detection result and the estimated sleep time information (S430).

Calculation of sleep time is important because AHI means the apnea-hypopnea index within sleep time. In the prior art, there is a problem that the AHI is not calculated accurately because the AHI is calculated by collectively calculating the sleep time from the bedtime start time to the wake-up time, and in the present invention, in order to solve this problem, as in the above step S420 AHI is calculated based on pure sleep time, so an accurate AHI can be calculated.

When the calculation of AHI is completed, OSA severity may be estimated and provided to the user (S440).

OSA severity is calculated based on the AHI value, and in general, OSA of less than 5 minutes per hour is normal, less than 5 to 15 minutes is mild, less than 15 to 30 minutes is moderate, and more than 30 minutes In this case, the degree of severity may be determined and provided to the user, but data may be provided to the user based on clinical results or an AHI index prepared for the user's sleep time.

Although not disclosed in FIGS. 4 to 7 of the present invention, additionally storing measurement data (ECG signal, feature value, OSA, pure sleep time, AHI, OSA severity) in a storage device, learning deep learning, medical institutions The method may further include transmitting the measurement data to a medical institution and planning a medical prescription or treatment by analyzing the measurement data at a medical institution.

In the present invention, including an embodiment, the structure and operation method of the sleep breathing state classification device are proposed, and in contrast to the prior art, the structure and operation method of the sleep breathing state classification device of the present invention use only a single ECG signal to detect obstructive It can diagnose apnea, so the size of the device can be drastically reduced.

In addition, by diagnosing obstructive apnea using deep learning, it is possible to provide a device and method for measuring obstructive apnea customized for each user.

In addition, the severity of obstructive apnea can be more accurately calculated by calculating the obstructive apnea-hypopnea index by separating the time of occurrence of obstructive apnea and actual sleep time.

In addition, a more accurate sleep time can be calculated by changing the measurement unit time of obstructive apnea.

Features, structures, effects, etc. described in the above embodiments are included in at least one embodiment of the present invention, and are not necessarily limited to only one embodiment. Furthermore, the features, structures, and effects illustrated in each embodiment can be combined or modified with respect to other embodiments by those skilled in the art in the field to which the embodiments belong.

Therefore, contents related to these combinations and variations should be construed as being included in the scope of the present invention. In addition, although the embodiments have been described above, these are only examples and do not limit the present invention, and those of ordinary skill in the field to which the present invention belongs can exemplify the above to the extent that does not deviate from the essential characteristics of the present embodiment. It will be seen that various variations and applications are possible that have not yet been made. For example, each component specifically shown in the embodiments can be modified and implemented. And differences related to these variations and applications should be construed as being included in the scope of the present invention as defined in the appended claims.

100: biosignal measuring unit
200: pre-processing unit 210: segmentation unit
220: labeling unit 230: filtering unit
300: feature extraction unit 310: peak detection unit
320: feature generator 330: resizing unit
400: diagnosis unit 410: OSA detection unit
420: sleep time estimation unit 430: AHI calculation unit
440: breathing state estimation unit

Claims (16)

a bio-signal measurement unit for measuring an ECG signal of a user while sleeping;
a pre-processing unit pre-processing the ECG signal;
a feature extraction unit extracting features of the preprocessed ECG signal; and
a diagnostic unit for diagnosing the user's respiratory state using the characteristics;
Including,
Sleep breathing state classification device, characterized in that the feature extraction unit and the diagnosis unit are composed of deep learning.
According to claim 1,
The pre-processing unit
a segmentation unit that subdivides the ECG signal into predetermined units using a discrete sliding window method;
a labeling unit for labeling each region of the ECG signal segmented in the segmentation unit; and
a filtering unit filtering the labeled labels;
Including,
The sleep breathing state classification apparatus, characterized in that the labeling unit performs labeling by grouping regions in front and behind of the region to be diagnosed by the diagnosis unit among each region subdivided in the segmentation unit.
According to claim 2,
Sleep breathing state classification device, characterized in that the re-diagnosis of the breathing state by resetting the subdivision unit of the subdivision unit to a unit smaller than the predetermined unit according to the reliability of the diagnosis unit.
According to claim 1,
The feature extraction unit
a peak detection unit detecting a peak of an R wave in the ECG signal;
a feature generating unit generating a feature based on the peak of the R wave; and
a resizing unit that resizes the feature;
Sleep breathing state classification device comprising a.
According to claim 4,
The characteristics are the interval from R wave to the next R wave in the electrocardiogram (RRI), ECG-Derived Respiration (EDR), and amplitude of ECG (AMP),
Wherein the resizing unit performs oversampling by the RRI and downsampling by the EDR using an interpolation method.
According to claim 1,
The diagnosis department
an OSA detection unit for detecting obstructive sleep apnea (OSA) using the above characteristics;
a sleep time estimator for estimating a pure sleep time using the feature;
an AHI calculation unit for calculating an apnea-hypopnea index (AHI) using the OSA and the sleep time; and
a respiratory state estimator estimating severity of sleep apnea based on the AHI;
Sleep breathing state classification device comprising a.
According to claim 6,
The deep learning is composed of a combination of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN),
The OSA and AHI are sleep breathing state classification apparatus, characterized in that for calculating the probability of each breath using softmax.
According to claim 7,
Sleep breathing state classification device, characterized in that for performing the learning of the convolutional neural network using the measurement result of the diagnostic unit and the ECG signal.
measuring an ECG signal in a sleep state;
Dividing the ECG signal into predetermined units;
extracting features of the divided ECG signal; and
diagnosing a respiratory state using the characteristics;
Including,
The step of extracting the feature and the step of diagnosing the breathing state are sleep breathing state classification method, characterized in that using a classification model composed of deep learning.
According to claim 9,
The step of dividing into the predetermined unit is
subdividing the ECG signal into predetermined units using a discrete sliding window method;
labeling each of the subdivided regions;
grouping labels before and after the label to be measured among the labeled bells and setting them as one group; and
filtering the group;
Sleep breathing state classification method comprising a.
According to claim 10,
Sleep breathing state classification method, characterized in that the re-diagnosis of the breathing state by subdividing the subdivided unit into smaller units than a predetermined unit according to the reliability of the breathing state.
According to claim 9,
Extracting the features
detecting a peak of an R wave in the divided ECG signal;
extracting a feature based on the peak of the R wave; and
resizing the feature;
Sleep breathing state classification method comprising a.
According to claim 12,
The characteristics are the interval from R wave to the next R wave in the electrocardiogram (RRI), ECG-Derived Respiration (EDR), and amplitude of ECG (AMP),
In the resizing step, the RRI performs oversampling and the EDR performs downsampling using an interpolation method.
According to claim 9,
Steps to diagnose breathing
Detecting obstructive sleep apnea (OSA) and pure sleep time using the feature;
Calculating an apnea-hypopnea index (AHI) based on the OSA and the sleep time; and
estimating the severity of sleep apnea;
Sleep breathing state classification method comprising a.
According to claim 14,
The deep learning is composed of a combination of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN),
The OSA and AHI are sleep breathing state classification method, characterized in that for calculating the probability of each breath using softmax.
According to claim 15,
The step of diagnosing the breathing state further comprises the step of performing learning of the convolutional neural network using the measurement result and the ECG signal.

Images