KR102068484B1 - Method for making prediction model for sleep apnea syndrome and method for predicting sleep apnea syndrome by using the same model - Google Patents

Abstract

Provided are a method for producing a model capable of predicting sleep apnea syndrome with high accuracy by using breathing sound or snoring sound of a sleeping person, and a method for predicting sleep apnea syndrome by using the model. The method for producing a prediction model for sleep apnea syndrome comprises the following steps of: preparing audio data obtained by recording breathing sound of each subject while a subject group consisting of a plurality of subjects is in a sleeping state; removing noise from the audio data; dividing the audio data into predetermined breathing cycle units, wherein each of the divided audio data is called a window; extracting an audio feature related to a feature of breathing sound for each window; and producing a prediction model for sleep apnea syndrome by performing machine learning on the subject group by using the audio feature of each subject and an apnea-hypopnea index as variables.

Description

Generation method for predicting sleep apnea syndrome and method for predicting sleep apnea syndrome by using the same model

The present invention relates to a method for generating a sleep apnea prediction model and a method for predicting sleep apnea using the model, and more specifically, a model capable of predicting sleep apnea with high accuracy using a breathing sound or snoring sound of a sleeping person. And a method for predicting sleep apnea using this model.

Sleep apnea is a case in which breathing stops for at least 10 seconds during sleep. Obstructive Sleep Apnea (OSA), which occurs due to obstruction of the oral cavity, despite attempts to breathe, Efforts to rest are classified as Central Sleep Apnea (CSA), which temporarily stops. Centralline sleep apnea is primarily associated with heart or brain problems, but obstructive sleep apnea is a symptom with snoring and is generally not treated as a serious illness. However, when obstructive sleep apnea occurs, the concentration of oxygen in the blood decreases, the concentration of carbon dioxide increases, and sympathetic nerve activation occurs due to repeated wakefulness, which may worsen hypertension, cardiovascular disease and cerebrovascular disease. Obstructive sleep apnea is present in about 25% of adult males, and is known to occur in more than 5% of the total population. It is known that the proportion increases as the obese population increases.

To date, the only way to know if you have obstructive sleep apnea is through polysomnography. Sleep polysomnography is a test that records and analyzes the state of sleep using various instruments to diagnose various abnormal conditions occurring during sleep. For example, electroencephalogram (EEG) for brain function status, safety test (EOG) for eye movement, electromyogram (EMG) for muscle condition, electrocardiogram (ECG) for heart rhythm, overall It's a way to test while taking a night's sleep. In particular, the process of diagnosing obstructive sleep apnea is a classical method of directly counting the number of times apnea or hypopnea occurs while the examinee sleeps. However, these polysomnography laboratories lack accessibility very much due to lack of numbers nationwide, and at least one night of waiting time is very long due to congestion. Therefore, it is practically difficult to prescribe a polysomnography to all patients with suspected obstructive sleep apnea.
<Preceding technical literature>
Publication No. 10-2016-0075677 (Publishing date: June 29, 2016)

The problem to be solved by the present invention is to provide a method for generating a sleep apnea prediction model that can predict sleep apnea with high accuracy through the machine learning using breathing sounds or snoring sounds.

Another object of the present invention is to provide a method for predicting sleep apnea using this model.

Problems to be solved by the present invention are not limited to the above-mentioned problems, and other problems not mentioned will be clearly understood by those skilled in the art from the following description.

A method of generating a sleep apnea prediction model according to an embodiment of the present invention for achieving the above object is to prepare audio data recording the breathing sound of each subject while the subject group consisting of a plurality of subjects in the sleep state step; Removing noise from the audio data; Dividing the audio data into predetermined respiratory cycle units (each of the divided audio data is called a window); Extracting audio features relating to features of the breath sound for each window; And generating a sleep apnea prediction model by performing machine learning on the target group using the audio feature and the apnea-low breathing index for each subject as variables.

After generating the sleep apnea prediction model, the method may further include verifying through 10-fold cross-validation.

The audio features include Beat Histogram, Area Method of Moments, Mel Frequency Cepstrum Coefficient (MFCC), Linear Predictive Coding, Area Method of Moments of ConstantQ-based MFCC, Area Method of Moments of Log of ConstantQ transform, Area Method of Moments of MFCC , And Method of Moments.

The apnea-low breathing index may be measured for each subject through a sleep polysomnography.

The breathing cycle may be 4 to 6 seconds.

The generating of the sleep apnea prediction model may include: specifying a target group as subjects having the apnea-low breathing index greater than or equal to a predetermined threshold, and referring to a plurality of different threshold values and a target group specified by each threshold value. After the individual sleep apnea prediction model is generated, the final sleep apnea prediction model can be selected by comparing the performance of each sleep apnea prediction model.

Sleep apnea prediction method according to an embodiment of the present invention for achieving the above another object, from the audio data of the breathing sound during sleep of the subject using the sleep apnea prediction model generated by the method to the subject Predict the apnea-low breathing index (AHI).

Specific details of other embodiments are included in the detailed description and drawings.

As described above, according to the method for generating a sleep apnea prediction model and the sleep apnea prediction method using the model, recording is performed through a mobile device while the examinee is sleeping at any place without having to visit the polysomnography room. Apnea-Hypopnea Index (AHI) can be easily predicted using a single breath or snoring sound. Specifically, since the audio feature defining the acoustic characteristics of apnea or hypopnea is extracted from the sleep respiration sounds recorded from a plurality of subjects, the apnea prediction model is generated by machine learning such as logistic regression analysis, and thus high accuracy. The apnea-low breathing index (AHI) can be predicted from the breathing sound of the subject.

1 is a diagram schematically illustrating a configuration of a system for predicting sleep apnea according to an embodiment of the present invention.
Figure 2 shows the information of the subject group used in the present experimental example.
3 shows the audio feature extracted in the present experimental example.
4 is a graph showing the performance of the sleep apnea prediction model according to the AHI threshold in the present experimental example.

Advantages and features of the present invention, and methods for achieving them will be apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various forms, and only the present embodiments make the disclosure of the present invention complete, and those of ordinary skill in the art may It is provided to fully convey the scope of the invention to those skilled in the art, and the present invention is defined only by the scope of the claims. Like reference numerals refer to like elements throughout.

1 is a diagram schematically illustrating a configuration of a system for predicting sleep apnea according to an embodiment of the present invention. The sleep apnea prediction system of the present invention generates a sleep apnea prediction model using the mobile device 10 for recording a subject's or subject's breathing sound, and the subject's recorded breathing sound and uses the model to determine whether the subject's sleep apnea is apnea. Sleep apnea prediction module for diagnosing the.

First, a target group consisting of a plurality of subjects is selected to generate a sleep apnea prediction model. For example, patients with Obstructive Sleep Apnea (OSA) or habitual snoring may be selected as a subject. If the subject has significant medical symptoms such as central sleep apnea (CSA), neurological disease, neuromuscular disease or heart failure, it is advisable to exclude them from the subject group. In the present invention, apnea (apnea) refers to a case in which breathing stops for at least 10 seconds during sleep, and hypopnea (hypopnea) does not completely stop breathing, but at least 10 seconds of breathing volume is reduced by 50% or more, or arousal or oxygen unsaturated ( <4%) associated with a slow decrease in respiratory volume for at least 10 seconds. Sleep apnea tests are performed on the subjects to determine the Apnea-Hypopnea Index (AHI). Apnea-low breathing index (AHI) is the average of total sleep time for 10 seconds of apnea or hypopnea without movement of the rib cage during sleep. Mild OSA for apnea-low respiratory index (AHI) of 5 or less than 15 times, moderate OSA for more than 15 and less than 30 times, and moderate OSA for more than 30 times. Can be classified.

The mobile device 10 is an electronic device which a user can carry and carry and includes a recording function, and is a device capable of transmitting and receiving data through a wired or wireless network with a peripheral device. For example, the mobile device 10 may be composed of an audio recorder, a video recorder, a smartphone, a mobile phone, a tablet PC, a wearable smart device, a PDA, a computer, a notebook, a digital TV, or the like. The mobile device 10 records the sleep breathing sound of the subject or the subject and converts the sound into audio data and transmits the sound to the sleep apnea prediction module 100. For example, sleep breath sounds can be converted to WAVE format audio data at 8kHz sampling frequency using the 'FFmpeg' software. In general, the sleep stage is composed of a transition to sleep (N1) stage, a light sleep stage (N2 stage), a deep sleep stage (N3 stage), a rapid eye movement stage (REM stage), and a wake stage. During the night, the sound of breathing occurring at all stages of sleep was recorded.

The sleep apnea prediction module 100 includes a noise removing unit 110, an audio feature extraction unit 120, a model generation and verification unit 130, a sleep apnea diagnosis unit 140, and a database 150. The database 150 stores various programs and data necessary for diagnosing sleep apnea.

The noise removing unit 110 removes noise from audio data received from the mobile device 10 as a preprocessing process. For example, the noise removing unit 110 may remove noise other than the breathing sound by a spectral subtraction method.

The audio feature extractor 120 divides the audio data recorded for one night into predetermined breathing cycle units. Here, each of the divided audio data is defined as a 'window', and the respiration period may be set to, for example, 4 to 6 seconds, preferably 5 seconds as an average inhalation and expiration period. The audio feature extractor 120 extracts an audio feature regarding the feature of the breathing sound for each window. Since the windows are divided into respiratory cycles, extracting audio features for each window is more advantageous for extracting acoustic features or patterns of apnea or hypopnea than for the entire audio data. For example, the audio feature extractor 120 may extract an audio feature by using an algorithm such as 'jAudio'.

In this embodiment, the audio features for specifying apnea or low breath are Beat Histogram, Area Method of Moments, Mel Frequency Cepstrum Coefficient (MFCC), Linear Predictive Coding, Area Method of Moments of ConstantQ-based MFCC, Area Method of Moments of Log of ConstantQ transform, Area Method of Moments of MFCC, and Method of Moments. Furthermore, audio features include Beat Sum, Compactness, Fraction Of Low Energy Windows, Peak Based Spectral Smoothness, Relative Difference Function, Root Mean Square, Spectral Centroid, Spectral Flux, Spectral Rolloff Point, Spectral Variability, Strength of Strongest Beat, Strongest Beat, Strongest Frequency Via Fast Fourier Transform (FFT) Maximum, Strongest Frequency Via Spectral Centroid, Strongest Frequency Via Zero Crossings, Zero Crossings, or combinations thereof may be further included.

The model generation and verification unit 130 uses the audio features extracted by the audio feature extraction unit 120 for each subject, and the apnea-low breathing index (AHI) measured through polysomnography for each subject as variables. , Or by performing machine learning, such as logistic regression, on all or part of the target group to generate a sleep apnea prediction model. In addition, the model generation and verification unit 130 performs a verification algorithm, such as cross-validation (eg, 10-fold cross-validation), leave-one-out, on the generated sleep apnea prediction model. 10-fold cross-validation is a method of dividing a data set into training data and test data and repeatedly performing modeling several times to obtain optimal parameters. First, divide the entire data set into 10 groups, then use 9 of them as a training set and the other 1 as a test set. Model parameters obtained after training in the training set are applied to the test set to measure the accuracy and performance of the model. This process is performed a total of 10 times using different groups as test sets, and the accuracy of the mean 10 times (mean cross-validated performance) is defined as the performance of the classifier. For example, the model generation and verification unit 130 may generate and verify a sleep apnea prediction model using an algorithm such as 'Weka'.

The model generation and verification unit 130 generates a prediction model for each threshold value for a target group having an apnea-low breathing index (AHI) equal to or greater than a plurality of threshold values, and compares the performance of each prediction model to an optimal prediction model. Can finally be selected. For example, the model generation and verification unit 130 generates a prediction model for a target group having an apnea-low breathing index (AHI) of 5 or more, 15 or more or 30, respectively, and performs the performance of each prediction model, for example. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy can be measured to compare each prediction model.

Using the sleep apnea prediction model generated as described above, the sleep apnea diagnosis unit 140 predicts an apnea-low breathing index (AHI) for the subject from the sound of breathing during the sleep of the subject.

Hereinafter , a method of generating a sleep apnea prediction model according to an embodiment of the present invention will be described in detail.

First, in order to generate a sleep apnea prediction model, the mobile device 10 records a sound of breathing during sleep of a plurality of subjects and converts the sound into audio data. While subjects sleep, polysomnography measures the apnea-low breathing index for each subject.

The noise removing unit 110 receives and prepares audio data from the mobile device 10 and then removes noise from the audio data.

The audio feature extractor 120 generates a plurality of windows by dividing the audio data into units of a predetermined breathing cycle. The breathing cycle here can be 4 to 6 seconds, preferably 5 seconds. The audio feature extractor 120 extracts an audio feature regarding the feature of the breathing sound for each window. Here audio features include Beat Histogram, Area Method of Moments, Mel Frequency Cepstrum Coefficient (MFCC), Linear Predictive Coding, Area Method of Moments of ConstantQ-based MFCC, Area Method of Moments of Log of ConstantQ transform, Area Method of Moments of MFCC , And the Method of Moments as the main configuration.

The model generation and verification unit 130 generates a sleep apnea prediction model by performing machine learning such as logistic regression analysis on the target group using audio features and apnea-low breathing index (AHI) for each subject as variables. do. The model generation and verification unit 130 performs a verification algorithm such as cross-validation (eg, 10-fold cross-validation), leave-one-out, etc. on the generated sleep apnea prediction model. The model generation and verification unit 130 specifies a target group consisting of subjects whose apnea-low breathing index (AHI) is greater than or equal to a predetermined threshold, generates an individual sleep apnea prediction model for a plurality of different thresholds, and The final sleep apnea prediction model is selected by comparing the performance of the sleep apnea prediction model.

Hereinafter , a method for predicting sleep apnea according to an embodiment of the present invention will be described in detail.

When the sleep apnea prediction model is generated based on the breathing sounds of the plurality of subjects through the above-described process, the sleep apnea diagnosis unit 140 uses the sleep apnea prediction model to determine the apnea-low breathing index for the subject ( AHI) is predicted.

Specifically, when the mobile device 10 records the breathing sound during sleep on the subject and converts the sound into audio data, the noise removing unit 110 removes the noise from the audio data, and the audio feature extracting unit 120 includes the audio. After dividing the data into the respiratory cycle unit, the audio feature is extracted for each window, and the sleep apnea diagnosis unit 140 inputs the audio feature of the subject into the sleep apnea prediction model to measure apnea-low breathing rate (AHI) for the subject. Predict.

Experimental Example

In this experiment, 116 patients (78 males, 38 females) were selected as subjects to generate a sleep apnea prediction model. Figure 2 shows the information of the subject group used in the present experimental example.

The mean age of the subjects was 50.4 years, the average body mass index was 25.5 Kg / m 2, and the average apnea-low breathing index (AHI) was 23 times / hr. The apnea-low breathing index (AHI) of each subject was measured by recording a breathing sound during sleep and performing a polysomnography. 28 patients with apnea-low breathing index (AHI) less than 5 (normal), 5 or less than 15 (mild OSA), 28 or more than 15 and less than 30 (moderate OSA), and 30 or more (severe OSA). It was confirmed by 30 people. FIG. 2 shows the number of subjects, average age, male and female ratio, average body mass index, and average AHI for each AHI interval.

The average total sleep time in this experimental example is 369.7 minutes. The average number of windows divided by 5 seconds of audio data of the breathing sound recorded during the total sleep period is 4436.8 per subject. The audio feature extractor 120 extracts a total of 508 audio features. 3 shows the audio feature extracted in the present experimental example.

Create apnea predictive model (Model 1) for subjects with an apnea-low breathing index (AHI) of 5 or higher, and predict a sleep apnea predictive model (Model 2) for subjects with an apnea-low breathing index (AHI) of 15 or higher. And apnea prediction model (Model 3) was generated for subjects with an apnea-low breathing index (AHI) of 30 or more. The performance of these predictive models, ie sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy, were measured. 4 is a graph showing the performance of the sleep apnea prediction model according to the AHI threshold in the present experimental example.

When the AHI thresholds for binary classification are defined as 5, 15, and 30, the number of subjects who accurately predicted the severity of obstructive sleep apnea was 96 in Model 1 (82.7%). Model 2 was identified as 98 (84.4%), and Model 3 was identified as 99 (85.3%).

When the threshold of AHI is 5 (ie, AHI is above 5), 88 subjects had mild to moderate OSA, the sensitivity of Model 1 was 87.5% (95% CI, 78.3-93.3), and the specificity was 67.8% (95% CI, 47.6-83.4), positive predictive value was 89.5% (95% CI, 80.6-94.8), and negative predictive value was 63.3% (95% CI, 46.9-79.5). The average AUC of the Receiver Operating Characteristic (ROC) curve was 0.83.

When the AHI threshold was 15 (ie, AHI was 15 or higher), 60 subjects had moderate to moderate OSA, the sensitivity of Model 2 was 81.6% (95% CI, 69.1-90.1), and the specificity was 87.5% (95% CI, 75.3-94.4), 87.5% (95% CI, 75.3-94.4) positive, and 81.6% (95% CI, 69.1-90.1) positive. The mean AUC of the ROC curve was determined to be 0.901.

At an AHI threshold of 30 (i.e. more than 30 AHI), 30 subjects had moderate OSA, the sensitivity of Model 3 was 60% (95% CI, 40.7-76.8), and the specificity was 94.1% (95 % CI, 86.3-97.8), positive predictive value was 78.2% (95% CI, 55.8-91.7), and negative predictive value was 87% (95% CI, 78.2-92.9). The mean AUC of the ROC curve was determined to be 0.91.

Among the three binary classifiers or the sleep apnea prediction models, the OSA prediction performance of Model 2 was the highest when the AHI threshold was 15.

In the present embodiment, the mobile device 10 and the sleep apnea prediction module 100 are expressed in separate configurations, but are not necessarily physically separated, but are functionally separated. For example, the entire configuration or some components of the sleep apnea prediction module 100 may be implemented in the form of an application installed in the mobile device 10. Specifically, the sleep apnea prediction module 100 may be implemented in one mobile application form and installed in the mobile device 10. In another example, some components of the sleep apnea prediction module 100 may be implemented in a mobile application behavior. Installed in the mobile device 10 and the rest of the configuration may be implemented in the form of an external server to be configured to communicate with the mobile device 10 via a wired or wireless network.

Although embodiments of the present invention have been described above with reference to the accompanying drawings, those skilled in the art to which the present invention pertains may implement the present invention in other specific forms without changing the technical spirit or essential features thereof. I can understand that. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive.

10: mobile device
100: sleep apnea prediction module
110: noise removing unit
120: audio feature extraction unit
130: model generation and verification unit
140: sleep apnea diagnosis unit
150: database

Claims (6)

Sleep Apnea Prediction Module:
Preparing audio data recording sound of each subject's breath while the subject group consisting of a plurality of subjects is in a sleeping state;
Removing noise from the audio data;
Dividing the audio data into predetermined respiratory cycle units (each of the divided audio data is called a window);
Extracting audio features relating to features of the breath sound for each window; And
And generating a sleep apnea prediction model by performing machine learning on the target group using the audio feature and the apnea-low breathing index for each subject as variables. The method of claim 1,
Generating the sleep apnea prediction model, and then verifying the result of the sleep apnea prediction model. The method of claim 1,
The audio features include Beat Histogram, Area Method of Moments, Mel Frequency Cepstrum Coefficient (MFCC), Linear Predictive Coding, Area Method of Moments of ConstantQ-based MFCC, Area Method of Moments of Log of ConstantQ transform, Area Method of Moments of MFCC A method of generating a sleep apnea predictive model, comprising:, and Method of Moments. The method of claim 1,
The apnea-low breathing index is a sleep apnea predictive model, characterized in that the measurement for each subject through a sleep test. The method of claim 1,
The respiration cycle is a method for generating a sleep apnea prediction model, characterized in that 4 to 6 seconds. The sleep apnea prediction module predicts an apnea-low apnea index (AHI) for the subject from the audio data of the breathing sound of the subject by using the sleep apnea prediction model generated by claim 1. Forecast method.

Images