Method for detecting sleep apnea and system therefor

Abstract

Provided are a method for detecting sleep apnea and a system therefor. The method for detecting sleep apnea according to some embodiments may comprise the steps of: acquiring an electrocardiogram (ECG) signal; configuring an input sequence of a deep learning model including a feature extractor and a recurrent neural network (RNN)-based predictor by preprocessing a plurality of segments of the ECG signal; extracting a feature sequence from the input sequence via the feature extractor; and detecting a segment in which sleep apnea occurs among the plurality of segments by inputting the feature sequence to the predictor. According to this method, a section in which sleep apnea occurs during a sleep period of a subject can be accurately detected.

Description

Sleep apnea detection method and system thereof

[0001] The present invention relates to a method for detecting sleep apnea and a system therefor.

[0002] Sleep apnea is a condition in which breathing stops for a certain period of time during sleep. Sleep apnea can be categorized into obstructive sleep apnea (OSA), which occurs when the oral airway is blocked despite attempts to breathe, and central sleep apnea (CSA), in which the breathing effort itself temporarily ceases. Of these, obstructive sleep apnea, which often occurs alongside snoring, is generally not considered a serious condition.

[0003] However, obstructive sleep apnea can cause chronic intermittent hypoxia, oxidative stress due to sleep disruption, and sympathetic nervous system activation due to repeated awakenings during sleep, which can ultimately lead to various physical harms. Research has shown that obstructive sleep apnea can cause or worsen diabetes, cardiovascular disease (e.g., hypertension), and cerebrovascular disease.

[0004] The diagnosis of sleep apnea is made through polysomnography. Polysomnography utilizes multiple sensors capable of detecting biological signals to record and analyze the subject's sleep state. However, polysomnography not only requires the subject's own sleeping space, but also requires various medical equipment and personnel. Furthermore, because polysomnography is performed in an unfamiliar environment while the subject wears numerous sensor devices, the accuracy of the test results frequently decreases (the so-called "first-night effect").

[0005] A technical problem to be solved through some embodiments of the present disclosure is to provide a method and system capable of accurately detecting sleep apnea in a subject.

[0006] Another technical challenge to be solved by some embodiments of the present disclosure is to provide a method and system for building a high-performance deep learning model for detecting sleep apnea.

[0007] Another technical challenge to be solved through some embodiments of the present disclosure is to provide a method and system for efficiently building a deep learning model for detecting sleep apnea.

[0008] The technical problems of the present disclosure are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art of the present disclosure from the description below.

[0009] In order to solve the above technical problem, a method for detecting sleep apnea according to some embodiments of the present disclosure is a method performed by at least one computing device, the method including: obtaining an electrocardiogram (ECG) signal; preprocessing a plurality of segments of the ECG signal to configure an input sequence of a deep learning model, wherein the deep learning model is configured to include a feature extractor and a recurrent neural network (RNN)-based predictor; extracting a feature sequence from the input sequence through the feature extractor; and inputting the feature sequence to the predictor to detect a segment in which sleep apnea has occurred among the plurality of segments.

[0010] In some embodiments, the feature extractor may be implemented based on a Convolutional Neural Network (CNN) that performs a 1D convolution operation.

[0011] In some embodiments, the step of configuring the input sequence may include the step of extracting a plurality of RRI (RR Interval) and RPA (R-Peak Amplitude) values ​​from a signal of a specific segment among the plurality of segments, and the step of generating data corresponding to the specific segment in the input sequence using the plurality of RRI and RPA values.

[0012] In some embodiments, the step of configuring the input sequence may include the steps of preprocessing a signal of a specific segment among the plurality of segments to generate first data, preprocessing a signal of an extended segment corresponding to the specific segment to generate second data, wherein the extended segment includes the specific segment and has a length longer than the specific segment, and determining a pair of the first data and the second data as data corresponding to the specific segment in the input sequence.

[0013] In some embodiments, the feature extractor comprises a first feature extractor, a second feature extractor, and a feature aggregator, wherein the second feature extractor is configured at least partially differently from the first feature extractor, and the step of extracting the feature sequence may include the steps of extracting a first feature from the first data via the first feature extractor, extracting a second feature from the second data via the second feature extractor, and aggregating the first feature and the second feature via the feature aggregator to generate an integrated feature, wherein the integrated feature is a feature corresponding to the specific segment in the feature sequence.

[0014] In some embodiments, the first feature extractor may be implemented with a deeper neural network than the second feature extractor.

[0015] In some embodiments, the first feature extractor may be implemented based on a ResNet (Residual neural network), and the second feature extractor may be implemented based on a CNN (Convolutional Neural Network) including an instance normalization layer.

[0016] In some embodiments, the step of generating the integrated feature may include the steps of combining the first feature and the second feature, analyzing a correlation between channels of the combined feature to derive a channel weight for the combined feature, and generating the integrated feature by reflecting the channel weight in the combined feature.

[0017] In some embodiments, the weight of a particular channel of the combined feature is derived by analyzing the correlation between the particular channel and neighboring channels, and the number of neighboring channels can be determined based on the total number of channels of the combined feature.

[0018] In some embodiments, the predictor comprises a first RNN block, a second RNN block, and a prediction layer, the feature sequence comprises a first feature corresponding to a first segment among the plurality of segments and a second feature corresponding to a second segment, and the step of predicting a segment in which sleep apnea occurs may include a step of inputting an output value of the first RNN block for the first feature to the prediction layer to predict whether sleep apnea occurs in the first segment, and a step of inputting an output value of the second RNN block for the second feature to the prediction layer to predict whether sleep apnea occurs in the second segment.

[0019] In some embodiments, the training process of the deep learning model may include: training the feature extractor by connecting a temporary predictor to the feature extractor, wherein the temporary predictor is a neural network configured to receive a feature of a specific segment extracted from the feature extractor and predict whether sleep apnea occurs in the specific segment; and performing additional training by connecting the RNN-based predictor to the trained feature extractor.

[0020] In order to solve the above-described technical problem, a sleep apnea detection system according to some embodiments of the present disclosure includes a memory storing one or more processors and a computer program executed by the one or more processors, wherein the computer program may include instructions for: obtaining an electrocardiogram (ECG) signal; preprocessing a plurality of segments of the ECG signal to configure an input sequence of a deep learning model, wherein the deep learning model includes a feature extractor and a recurrent neural network (RNN)-based predictor; extracting a feature sequence from the input sequence through the feature extractor; and inputting the feature sequence to the predictor to detect a segment in which sleep apnea has occurred among the plurality of segments.

[0021] In order to solve the above-described technical problem, a computer program according to some embodiments of the present disclosure may be stored in a computer-readable recording medium to execute the steps of: obtaining an electrocardiogram (ECG) signal; preprocessing a plurality of segments of the ECG signal to configure an input sequence of a deep learning model, wherein the deep learning model includes a feature extractor and a recurrent neural network (RNN)-based predictor; extracting a feature sequence from the input sequence through the feature extractor; and inputting the feature sequence into the predictor to detect a segment in which sleep apnea has occurred among the plurality of segments.

[0022] According to some embodiments of the present disclosure, a first feature (i.e., a local / deep feature) can be extracted by performing an in-depth analysis on a basic segment of an electrocardiogram signal through a feature extractor of a deep learning model, and a second feature (i.e., a global / shallow feature) can be extracted by performing a shallow analysis on an extended segment. Then, whether sleep apnea occurs or not can be predicted from an integrated feature that is an aggregate of the first and second features through a predictor of the deep learning model. In this case, a sleep apnea occurrence section can be accurately detected (diagnosed) from an electrocardiogram signal measured during a subject's sleep period (time). That is, by comprehensively considering feature information at various levels and predicting whether sleep apnea occurs for each segment of the electrocardiogram signal, the sleep apnea occurrence section can be accurately detected.

[0023] Additionally, the feature extractor of a deep learning model may be configured to include a first feature extractor and a second feature extractor, with the first feature extractor implemented as a deeper neural network than the second feature extractor. In this case, local / deep features can be accurately extracted from the base segment (i.e., a narrow section of data), and global / shallow features can be accurately extracted from the extended segment (i.e., a wide section of data).

[0024] In addition, since the first feature extractor and / or the second feature extractor are configured to extract features through a 1D convolution operation, contextual information of the input data can be effectively summarized during feature extraction.

[0025] In addition, since the feature aggregator is configured to aggregate the first feature and the second feature based on the channel weight, the feature values ​​of important channels can be emphasized during feature aggregation, and as a result, the accuracy of sleep apnea detection can be further improved.

[0026] Additionally, the deep learning model's predictor can be implemented based on a Recurrent Neural Network (RNN) and configured to receive feature sequences for different (e.g., consecutive) segments (i.e., basic segments) and predict the presence or absence of sleep apnea for each segment. In this case, the presence or absence of sleep apnea in a specific segment can be predicted by considering information from surrounding segments, further improving the accuracy of sleep apnea detection.

[0027] Furthermore, a deep learning model for sleep apnea detection can be built by first training a feature extractor by connecting a lightweight temporary predictor, and then performing additional training by connecting the original predictor to the trained feature extractor. In this case, compared to training the deep learning model all at once, parameters such as the feature extractor and predictor can be updated more quickly, significantly reducing the time and computing costs required for deep learning model training.

[0028] The effects according to the technical idea of ​​the present disclosure are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

[0029] FIG. 1 is an exemplary diagram schematically illustrating the operation of a sleep apnea detection system according to some embodiments of the present disclosure.

[0030] FIG. 2 is an exemplary flowchart schematically illustrating a sleep apnea detection method according to some embodiments of the present disclosure.

[0031] FIG. 3 is an exemplary diagram illustrating the structure and operation of a deep learning model according to some embodiments of the present disclosure.

[0032] FIG. 4 is an exemplary diagram for explaining the structure and operation of a feature extractor of a deep learning model according to some embodiments of the present disclosure.

[0033] FIG. 5 is an exemplary drawing for explaining the detailed structure and operation of a first feature extractor according to some embodiments of the present disclosure.

[0034] FIG. 6 is an exemplary drawing for explaining the detailed structure and operation of a second feature extractor according to some embodiments of the present disclosure.

[0035] FIG. 7 is an exemplary diagram illustrating the detailed structure and operation of a feature aggregator according to some embodiments of the present disclosure.

[0036] FIG. 8 and FIG. 9 are exemplary drawings for explaining the structure and operation of a predictor of a deep learning model according to some embodiments of the present disclosure.

[0037] Figure 10 is an exemplary flowchart showing the detailed process of the deep learning model learning step illustrated in Figure 2.

[0038] Figure 11 is an exemplary flowchart showing a detailed process of the input sequence configuration step illustrated in Figure 10.

[0039] FIG. 12 and FIG. 13 are exemplary drawings for explaining a segment preprocessing method according to some embodiments of the present disclosure.

[0040] FIG. 14 is an exemplary diagram illustrating a method for learning a deep learning model according to some embodiments of the present disclosure.

[0041] FIG. 15 illustrates an exemplary computing device that may implement a sleep apnea detection system according to some embodiments of the present disclosure.

[0042] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings. The advantages and features of the present invention, and methods for achieving them, will become clearer with reference to the embodiments described in detail below together with the attached drawings. However, the technical spirit of the present invention is not limited to the following embodiments, but can be implemented in various different forms. The following embodiments are provided only to complete the technical spirit of the present invention and to fully inform those skilled in the art of the present invention of the scope of the present invention, and the technical spirit of the present invention is defined only by the scope of the claims.

[0043] When assigning reference numerals to components in each drawing, it should be noted that identical components are assigned the same numerals whenever possible, even if they appear on different drawings. Furthermore, when describing the present invention, if a detailed description of a related known configuration or function is deemed likely to obscure the gist of the present invention, such detailed description will be omitted.

[0044] Unless otherwise defined, all terms (including technical and scientific terms) used herein may be used in their common sense to those of ordinary skill in the art to which the present invention pertains. Furthermore, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise. The terminology used herein is for the purpose of describing embodiments and is not intended to limit the present invention. In this specification, singular forms also include plural forms, unless specifically stated otherwise.

[0045] Additionally, terms such as first, second, A, B, (a), (b), etc. may be used to describe components of the present invention. These terms are only intended to distinguish the components from other components, and the nature, order, or sequence of the components are not limited by the terms. When it is described that a component is "connected," "coupled," or "connected" to another component, it should be understood that the component may be directly connected or connected to the other component, but another component may also be "connected," "coupled," or "connected" between each component.

[0046] As used herein, the terms “comprises” and / or “comprising” do not exclude the presence or addition of one or more other components, steps, operations and / or elements.

[0047] Hereinafter, various embodiments of the present invention will be described in detail with reference to the attached drawings.

[0048] FIG. 1 is an exemplary drawing schematically illustrating the operation of a sleep apnea detection system (10) according to some embodiments of the present invention.

[0049] As illustrated in Fig. 1, the sleep apnea detection system (10) is a computing device / system that can detect (diagnose) sleep apnea (SA) by analyzing the subject's electrocardiogram ('ECG') signal (12). For example, the sleep apnea detection system (10) can predict (diagnose) whether sleep apnea has occurred (see 'SA', 'NSA (Non-SA)') or detect the sleep apnea occurrence section (i.e., the time section corresponding to the segment in which sleep apnea has occurred in the ECG signal (12)) by analyzing the subject's electrocardiogram ('ECG') signal (12) measured during the sleep time (period) through a deep learning model (11). Hereinafter, for convenience of explanation, the sleep apnea detection system (10) will be abbreviated as a 'detection system'.

[0050] The specific method by which the sleep apnea detection system (10) builds (learns) a deep learning model (11) and detects sleep apnea through this will be described in detail later with reference to drawings below FIG. 2.

[0051] For reference, the reason why sleep apnea is detected by analyzing electrocardiogram signals (e.g., 12) is because breathing and cardiac activity are closely related. To elaborate, since changes in breathing are accompanied by changes in cardiac activity (i.e., when apnea occurs during sleep, sympathetic nerve activity increases, and when apnea resolves, parasympathetic nerve activity increases, and these changes in sympathetic and parasympathetic nerve activity affect cardiac activity), the breathing state is reflected in the electrocardiogram signal, which records cardiac activity. Therefore, analyzing the electrocardiogram signal can accurately determine the subject's breathing state, and for this reason, it can be understood that analyzing the electrocardiogram signal is used to detect sleep apnea.

[0052] Sleep apnea may sometimes be referred to as a "sleep disorder." The term "sleep apnea" can also encompass the concept of sleep hypopnea.

[0053] In addition, since the deep learning model (11) is a model used for sleep apnea detection (diagnosis or prediction), it may be named as a 'sleep apnea (symptom) detection (diagnosis or prediction) model' or a 'detection (diagnosis or prediction) model' depending on the case.

[0054] Additionally, the detection system (10) may also be named as ‘sleep apnea diagnosis system (module or device)’, ‘sleep apnea prediction system (module or device)’, etc.

[0055] The detection system (10) described above may be implemented in at least one computing device. For example, the detection system (10) may be implemented in a single computing device. The first function of the detection system (10) may be implemented in a first computing device, and the second function may be implemented in a second computing device. Alternatively, specific functions of the detection system (10) may be implemented in multiple computing devices.

[0056] A computing device may include any device equipped with computing (processing) capabilities. For an example of such a device, see FIG. 15. Since a computing device is a collection of interacting components (e.g., a processor, memory, etc.), it may sometimes be referred to as a "computing system." Of course, the term "computing system" can also encompass the concept of a collection of interacting computing devices.

[0057] So far, the operation of the detection system (10) according to some embodiments of the present disclosure has been briefly described with reference to FIG. 1. Hereinafter, various methods that can be performed in the detection system (10) described above will be described in detail with reference to the drawings including FIG. 2 and below.

[0058] For ease of understanding, the following description assumes that all steps / operations of the methods described below are performed on the aforementioned detection system (10). Therefore, if the subject of a specific step / operation is omitted, it can be understood that it is performed on the detection system (10). Of course, in a real-world environment, some steps / operations of the methods described below may be performed on other computing devices. For example, the training of a deep learning model (11) may be performed on other computing devices, depending on the circumstances.

[0059] Figure 2 is an exemplary flowchart schematically illustrating a sleep apnea detection method according to some embodiments of the present disclosure. However, this is merely a preferred embodiment for achieving the purpose of the present disclosure, and some steps may be added or deleted as needed.

[0060] As illustrated in FIG. 2, the sleep apnea detection method according to the embodiments may begin at step S21, in which a deep learning model (11) is trained using an electrocardiogram signal dataset (i.e., a training set) of a sleep apnea patient. The electrocardiogram signal dataset may be composed of, for example, single-lead electrocardiogram signal samples, but the scope of the present disclosure is not limited thereto.

[0061] Hereinafter, the detailed structure and learning process of the deep learning model (11) will be described in detail with reference to FIGS. 3 to 13.

[0062] First, for convenience of understanding, the structure and operating principles of a deep learning model (11) according to some embodiments of the present disclosure will be described with reference to FIGS. 3 to 9.

[0063] FIG. 3 is an exemplary drawing for explaining the structure and operation of a deep learning model (11) according to some embodiments of the present disclosure.

[0064] As illustrated in FIG. 3, a deep learning model (11) according to embodiments may be configured to include a feature extractor (31) and a predictor (32).

[0065] The feature extractor (31) is a module that extracts features for segments (e.g., 33, 34) of an electrocardiogram signal. As illustrated, the feature extractor (31) receives data (35, 36) extracted from each of the basic segment (33) and its extended segment (34) of the electrocardiogram signal, analyzes them, and extracts features.

[0066] Here, the basic segment (33) refers to a segment of a basic unit, and the extended segment (34) may refer to a segment having a longer length (i.e., time length) than the basic segment (33). The extended segment (34) may include a corresponding basic segment (33). Fig. 3 assumes that the length of the extended segment (34) is '5 times' that of the basic segment (33) (e.g., the length of the basic segment (33) is '1 minute' and the length of the extended segment (34) is '5 minutes'). The basic segment (e.g., 33) may also be abbreviated as 'segment' in some cases.

[0067] Hereinafter, for the clarity of the present disclosure, data (eg, 35) extracted from a basic segment (eg, 33) is named 'first data', and data (eg, 36) extracted from an extended segment (eg, 34) is named 'second data'.

[0068] As illustrated in FIG. 4, the feature extractor (31) may be configured to include a first feature extractor (41), a second feature extractor (42), and a feature aggregator (43).

[0069] The first feature extractor (41) is a module that extracts a feature (44, hereinafter referred to as the 'first feature') from the first data (35). The first feature (44) can be understood as a feature extracted through a more in-depth analysis than, for example, the second feature (45). That is, the first feature (44) can be understood as a feature extracted through an in-depth analysis of the data (35) of the basic segment (33) (i.e., data of a narrower section), and may be named as a 'local feature', a 'deep feature', etc. in some cases. In addition, the feature (e.g., 44, 45) may be named as a 'feature map', 'feature data', a 'feature vector', etc. in some cases.

[0070] The first feature extractor (41) can be implemented (configured) based on various types of neural networks, and the specific implementation method thereof may vary depending on the embodiment.

[0071] In some embodiments, the first feature extractor (41) may be implemented (configured) based on a Convolutional Neural Network (CNN, i.e., '1D CNN') that performs a 1D convolution operation. For example, the first feature extractor (41) may be configured to include layers that perform a 1D convolution operation (i.e., '1D convolution layers'). By doing so, the first feature (44) may be extracted by taking into account the contextual characteristics of the first data (35).

[0072] In some other embodiments, the first feature extractor (41) may be implemented (configured) as a deeper neural network than the second feature extractor (42). For example, the first feature extractor (41) may be configured to include a greater number of layers (or blocks) than the second feature extractor (42). By doing so, in-depth analysis of the first data (35) can be accurately performed. As a specific example, the first feature extractor (41) may be implemented based on a ResNet (Residual Neural Network) (see FIG. 5). ResNet is a neural network composed of a plurality of residual blocks, and is a neural network designed to easily increase depth by using skip connections (i.e., a deep neural network that solves the gradient vanishing problem). Those skilled in the art will likely already be familiar with the detailed structure and operating principles of ResNet, and therefore, a detailed description thereof will be omitted.

[0073] In some other embodiments, the first feature extractor (41) may be implemented (configured) based on various combinations of the above-described embodiments. For example, as illustrated in FIG. 5, the first feature extractor (41) may be implemented based on a 1D CNN and a ResNet. Specifically, the first feature extractor (41) may be implemented based on a combination of a plurality (e.g., 'N') of residual blocks (51) and a 1D convolutional layer. FIG. 5 illustrates a case where the residual block (51) is also composed of 1D convolutional layers, and a 'ReLU (Rectified Linear Unit) function' is used as an activation function of the 1D convolutional layer.

[0074] This is explained again with reference to Figure 4.

[0075] The second feature extractor (42) is a module that extracts a feature (45, hereinafter referred to as the “second feature”) from the second data (36). The second feature (45) can be understood as a feature extracted through a shallow analysis compared to the first feature (44), for example. That is, the second feature (45) can be understood as a feature extracted through a comprehensive analysis of the data (36) of the extended segment (34) (i.e., data of a wider section), and may be named as a “global feature”, a “shallow feature”, etc. depending on the case. The second feature (45) can be understood as having a role of transmitting feature information obtained by comprehensively analyzing signals surrounding the basic segment (33) to the predictor (32).

[0076] The second feature extractor (42) can also be implemented (configured) based on various types of neural networks, and the specific implementation method thereof may vary depending on the embodiment.

[0077] In some embodiments, the second feature extractor (42) may also be implemented (configured) based on a 1D CNN. For example, the second feature extractor (42) may also be configured to include 1D convolutional layers. By doing so, the second feature (45) may be extracted by taking into account the time-series characteristics of the second data (36).

[0078] In some other embodiments, the second feature extractor (42) may be implemented (configured) as a shallower neural network than the first feature extractor (41). That is, the second feature extractor (42) may be configured to include fewer layers (or blocks) than the first feature extractor (41). By doing so, shallow (comprehensive) analysis of the second data (36) can be accurately performed.

[0079] In some other embodiments, the second feature extractor (42) may be implemented (configured) based on various combinations of the above-described embodiments. For example, as illustrated in FIG. 6, the second feature extractor (42) may be implemented as a neural network based on a 1D CNN but shallower than the first feature extractor (41). FIG. 6 illustrates a case where the second feature extractor (42) is configured to include a 1D convolution layer and an instance normalization layer, and is configured to output a second feature (45) of the same size as the first feature (44) through a max pooling operation or the like. In addition, FIG. 6 illustrates a case where a ReLU function is applied after the instance normalization operation, and a dropout technique is applied to some layers. Those working in the relevant technical field will likely already be familiar with the concept and operating principles of instance normalization, which is performed on an instance basis rather than a batch basis, so an explanation of this will be omitted.

[0080] This is explained again with reference to Figure 4.

[0081] The feature aggregator (43) is a module that aggregates the first feature (44) and the second feature (45) (see 46). Hereinafter, for the clarity of the present disclosure, the feature generated by aggregating the first feature (44) and the second feature (45) will be referred to as an 'integrated feature (46)'.

[0082] The feature aggregator (43) can be implemented (configured) in various ways.

[0083] In some embodiments, the feature aggregator (43) may be implemented as a module that performs an aggregating operation (e.g., concatenation, elementwise multiplication, addition, averaging, etc.). For example, the feature aggregator (42) may be configured to aggregate the first feature (44) and the second feature (45) through a concatenation operation.

[0084] In some other embodiments, the feature aggregator (43) may be configured to aggregate the first feature (44) and the second feature (45) via a neural network module / layer, such as a Multi-Layer Perceptron (MLP).

[0085] In some other embodiments, the feature aggregator (43) may be configured to aggregate the first feature (44) and the second feature (45) using the channel-specific weights derived through the channel attention operation. For example, as illustrated in FIG. 7, the feature aggregator (43) may combine the first feature (44) and the second feature (45) through a concatenation operation or the like, and extract the channel representative values ​​(72) of the combined feature (71) through a global average pooling ('GAP') operation or the like. Next, the feature aggregator (43) may analyze the inter-channel correlation based on the channel representative values ​​(72) to derive the channel weights (73, i.e., 128 channel-specific weights) for the combined feature (71). At this time, the feature aggregator (43) may derive channel weights (73) by performing a 1D convolution operation using a kernel (so-called 'adaptive kernel') having an adaptive size (i.e., performing a 1D convolution operation in the channel direction), but the scope of the present disclosure is not limited thereto. The channel weights may also be named 'channel attention (value)' in some cases. Next, the feature aggregator (43) may generate an integrated feature (46) by reflecting (e.g., multiplying) the channel weights (73) to the combined feature (71). By doing so, the feature values ​​of important channels can be emphasized during feature aggregation, and as a result, the sleep apnea detection accuracy of the deep learning model (11) can be improved.

[0086] In the above embodiments, the size of the adaptive kernel (filter) may be determined based on the total number of channels of the combined feature (71). Here, the size of the adaptive kernel may be understood to mean the number of neighboring channels to be analyzed for correlation with a specific channel. For example, the size of the adaptive kernel may be determined according to the following mathematical expression 1. In the following mathematical expression 1, 'k' represents the size of the kernel, '|x|_odd' represents the odd number closest to x, and 'c' represents the total number of channels of the combined feature (71). In addition, 'γ' and 'g' represent fitting factors, which are constant values ​​for appropriately adjusting the value of 'k'. According to the following mathematical expression 1, as the number of channels of the feature (e.g., 71) increases, the size (length) of the adaptive kernel also increases (lengthens) accordingly.

[0087]

[0088] In some other embodiments, the feature aggregator (43) may be implemented (configured) based on various combinations of the embodiments described above.

[0089] Meanwhile, FIG. 4 illustrates a case where two features (44, 45) are extracted through two feature extractors (41, 42), but in some cases, three or more features may be extracted by using an additional feature extractor (not shown). For example, the feature extractor (31) may further include a third feature extractor (not shown) that extracts a medium feature (e.g., a feature extracted from a medium-length segment that is longer than the basic segment (33) and shorter than the extended segment (34).

[0090] This is explained again with reference to Figure 3.

[0091] The predictor (32) is a module that predicts whether sleep apnea occurs (or the degree of sleep apnea, etc.) based on the features output from the feature extractor (41) (i.e., the integrated features (46) of FIG. 4). For example, as illustrated, the predictor (32) may be configured to perform binary classification regarding the occurrence of sleep apnea, but the scope of the present disclosure is not limited thereto. In the following, the description will continue on the assumption that the predictor (32) is configured to predict the occurrence of sleep apnea.

[0092] The predictor (32) can be implemented (configured) based on various types of neural networks, and the specific implementation method may vary depending on the embodiment.

[0093] In some embodiments, the predictor (32) may be implemented (configured) based on a fully-connected layer. For example, the predictor (32) may be configured to include one or more fully-connected layers and receive an integrated feature (46) as input to predict whether sleep apnea occurs for a given segment (i.e., a base segment).

[0094] In some other embodiments, as illustrated in FIG. 8, the predictor (32) may be implemented based on a Recurrent Neural Network (RNN). Specifically, the predictor (32) may be configured to include a plurality of RNN blocks (82-1 to 82-K) and a prediction layer (83). The RNN blocks (82-1 to 82-K) may be configured to receive feature sequences (84-1 to 84-K), analyze them in consideration of the order, and output the analysis results. At this time, each feature constituting the feature sequences (84-1 to 84-K) means an integrated feature (e.g., 46) corresponding to different (e.g., consecutive) basic segments. The prediction layer (83) may be configured to receive the output values ​​of each of the RNN blocks (82-1 to 82-K) and predict whether sleep apnea occurs in the corresponding basic segment. For example, if the first RNN block (e.g., 82-1) receives a first integrated feature (e.g., 84-1) corresponding to the first basic segment as input and outputs a first output value, the prediction layer (83) can receive the first output value and predict whether sleep apnea occurs in the first basic segment (e.g., see 85-1). In addition, if the second RNN block (e.g., 82-2) receives a second integrated feature (e.g., 84-2) corresponding to the second basic segment (e.g., a basic segment temporally positioned after the first basic segment) and outputs a second output value, the prediction layer (83) can receive the second output value and predict whether sleep apnea occurs in the second basic segment (e.g., see 85-2). By doing so, whether sleep apnea occurs can be accurately predicted for each basic segment (i.e., the sleep apnea occurrence section can be accurately detected). The prediction layer (83) can be implemented (configured) based on, for example, a fully connected layer.

[0095] In the above embodiments, the RNN block (e.g., 82-1) may be implemented based on a Gated Recurrent Unit (GRU) cell, for example, as illustrated in FIG. 9. However, the scope of the present disclosure is not limited thereto. The RNN blocks (82-1 to 82-K) correspond to modules that comprehensively analyze features (i.e., feature sequences) corresponding to different segments, and thus may be named as a "feature analyzer", a "feature sequence analyzer", etc., as the case may be.

[0096] In some other embodiments, the predictor (32) may be implemented based on a neural network of the transformer series. For example, the predictor (32) may be configured to include a transformer and a prediction layer (e.g., 83). The transformer can input a feature sequence, comprehensively analyze it, and output an output value corresponding to each basic segment. Similarly to the above, the prediction layer (e.g., 83) can predict whether sleep apnea occurs for each basic segment.

[0097] So far, the structure and operating principles of the deep learning model (11) according to some embodiments of the present disclosure have been described with reference to FIGS. 3 to 9.

[0098] Hereinafter, the learning process / method of the deep learning model (11) described above will be described in detail with reference to FIGS. 10 to 14. The drawings below FIG. 10 assume that the predictor (32) of the deep learning model (11) is implemented based on 'RNN'.

[0099] Figure 10 is an exemplary flowchart illustrating the detailed process of step S21 described above (i.e., the training process of the deep learning model (11)). However, this is merely a preferred embodiment for achieving the purpose of the present disclosure, and some steps may be added or deleted as needed.

[0100] In step S101, multiple segments of electrocardiogram signal samples may be preprocessed to form an input sequence of a deep learning model (11). The detailed process of this step is illustrated in Fig. 11. The following description will be given with reference to Fig. 11.

[0101] It can be understood that steps S111 to S114, which will be described later, are performed on a plurality of basic segments and extended segments to form an input sequence.

[0102] In step S111, a basic segment and an extended segment corresponding thereto may be extracted from an electrocardiogram signal sample. For example, as illustrated in FIG. 12, the detection system (10) may extract basic segments (e.g., 123, 124, etc.) by moving a basic window (121) (i.e., through a sliding window technique), and may extract extended segments (e.g., 126) corresponding to each of the basic segments (e.g., 123, 124, etc.) by moving an extended window (125). However, the scope of the present disclosure is not limited thereto.

[0103] In step S112, the signal of the basic segment may be preprocessed to generate first data. For example, the detection system (10) may extract a predetermined number of RRI (RR Interval) and RPA (R Peak Amplitude) values ​​from the signal of the basic segment to generate the first data. As illustrated in FIG. 13, RRI refers to the interval between R peaks in an electrocardiogram signal, and RPA refers to the amplitude of an R peak. Since these contents correspond to matters self-evident to those skilled in the art, further explanation will be omitted.

[0104] The first data may be composed of, for example, a predetermined number of (RRI, RPA) tuples (i.e., the first data also corresponds to time series / sequence data). Drawings such as FIG. 3 assume that the first input data (35) consists of 180 (RRI, RPA) tuples. The detection system (10) may extract a predetermined number of (RRI, RPA) tuples for each basic segment using an interpolation method (e.g., cubic interpolation), but the scope of the present disclosure is not limited thereto.

[0105] In step S113, the signal of the extended segment may be preprocessed to generate second data. For example, the detection system (10) may perform preprocessing in the same manner as the basic segment to generate the second data. Drawings such as FIG. 3 assume that the second data (36) consists of 900 (RRI, RPA) tuples.

[0106] In step S114, a pair of first data and second data may be determined as data of an input sequence corresponding to the corresponding basic segment (i.e., a data element of the sequence). For example, referring again to FIG. 3, it may be understood that a pair of first data (35) and second data (36) constitutes a data element corresponding to a specific basic segment (33) in the input sequence.

[0107] The detection system (10) can configure an input sequence of the deep learning model (11) by repeatedly performing the above-described steps S111 to S114 for other basic segments and extended segments. The input sequence is composed of data elements corresponding to a plurality of basic segments, for example, and each of the data elements can be composed of a pair of first data extracted from a basic segment and second data extracted from a corresponding extended segment.

[0108] This is explained again with reference to Figure 10.

[0109] In step S102, the input sequence can be input into the deep learning model (11) to predict whether sleep apnea occurs for each segment (i.e., basic segment). For example, the detection system (10) can input (feed) each data element of the input sequence into the feature extractor (31) to extract an integrated feature for each data element. The multiple integrated features extracted in this way constitute a feature sequence (e.g., 84-1 to 84-K of FIG. 8). Next, the detection system (10) can input the feature sequence into the predictor (32) to predict whether sleep apnea occurs for each basic segment. For this, refer to the description of FIG. 8 for further details.

[0110] In step S103, the parameters of the deep learning model (11) may be updated based on the difference (i.e., loss) between the prediction result and the correct label. For example, the detection system (10) may update the parameters of the feature extractor (31) and the predictor (32) based on the loss (e.g., cross-entropy loss) between the prediction result for each basic segment obtained through the predictor (32) and the correct label.

[0111] Steps S101 to S103 described above can be repeatedly performed for other electrocardiogram signal samples included in the electrocardiogram signal dataset (i.e., training set). By doing so, the deep learning model (11) can be equipped with the ability to accurately predict (diagnose) the occurrence of sleep apnea (or accurately detect the sleep apnea occurrence period).

[0112] This is explained again with reference to Figure 2.

[0113] In step S22, an electrocardiogram signal of the subject (e.g., a single-lead electrocardiogram signal) can be acquired. Here, the electrocardiogram signal refers to a signal measured during the subject's sleep time (period).

[0114] In step S23, the subject's electrocardiogram signal can be analyzed through the learned deep learning model (11) to detect the section in which sleep apnea occurred. Specifically, the detection system (10) preprocesses the electrocardiogram signal in the same manner as described above to configure an input sequence, and inputs this into the learned deep learning model (11) to predict whether sleep apnea occurred for each basic segment. Based on the prediction result, the detection system (10) can detect the segment in which sleep apnea occurred among the basic segments of the electrocardiogram signal (i.e., the section in which sleep apnea occurred during the sleep period).

[0115] Hereinafter, a sleep apnea detection method according to some embodiments of the present disclosure has been described with reference to FIGS. 2 to 13. As described above, a first feature (i.e., a local / deep feature) can be extracted by performing an in-depth analysis on a basic segment of an electrocardiogram signal through a feature extractor (31) of a deep learning model (11), and a second feature (i.e., a global / shallow feature) can be extracted by performing a shallow analysis on an extended segment. In addition, whether or not sleep apnea occurs can be predicted from an integrated feature that is an aggregate of the first and second features through a predictor (32) of the deep learning model (11). In this case, a sleep apnea occurrence section can be accurately detected (diagnosed) from an electrocardiogram signal measured during a subject's sleep period (time). That is, by comprehensively considering feature information at various levels and predicting whether or not sleep apnea occurs for each segment of the electrocardiogram signal, a sleep apnea occurrence section can be accurately detected.

[0116] In addition, the feature extractor (31) of the deep learning model (11) may be configured to include a first feature extractor (41) and a second feature extractor (42), and the first feature extractor (41) may be implemented as a neural network deeper than the second feature extractor (42). In this case, local / deep features can be accurately extracted from the basic segment (i.e., data in a narrow section), and global / shallow features can be accurately extracted from the extended segment (i.e., data in a wide section).

[0117] In addition, since the first feature extractor (41) and / or the second feature extractor (42) are configured to extract features through a 1D convolution operation, the time series characteristics of the input data can be sufficiently considered when extracting features.

[0118] In addition, since the feature aggregator (43) is configured to aggregate the first feature and the second feature based on the channel weight, the feature values ​​of important channels can be emphasized during feature aggregation, and as a result, the accuracy of sleep apnea detection can be further improved.

[0119] In addition, the predictor (32) of the deep learning model (11) can be implemented based on RNN and configured to receive feature sequences for different (e.g., consecutive) segments (i.e., basic segments) and predict whether sleep apnea occurs for each segment. In this case, whether sleep apnea occurs in a specific segment can be predicted by considering information from surrounding segments, thereby further improving the accuracy of sleep apnea detection.

[0120] Below, a learning method for more efficiently constructing a deep learning model (11) will be described with reference to Fig. 14.

[0121] FIG. 14 is an exemplary diagram illustrating a method for learning a deep learning model according to some embodiments of the present disclosure.

[0122] As illustrated in Fig. 14, the detection system (10) can train a deep learning model (11) through a first learning stage and a second learning stage. The first learning stage and the second learning stage may be named, in some cases, as a "pretraining stage" and an "additional learning (or fine-tuning) stage."

[0123] Specifically, in the first learning stage, the detection system (10) can perform learning by connecting a temporary predictor (141) to the feature extractor (31). For example, the detection system (10) can predict whether sleep apnea occurs in a specific segment (see 142) of an electrocardiogram signal sample through the feature extractor (31) and the temporary predictor (141), and update the parameters of the feature extractor (31) and the temporary predictor (141) based on the loss (145) between the prediction result (143) and the correct label (144).

[0124] The temporary predictor (141) may be a neural network having fewer parameters than the predictor (32), and may be implemented based on a fully connected layer (FC) as illustrated. By training the feature extractor (31) using a neural network (i.e., the temporary predictor (141)) that is lighter than the predictor (32), training of the feature extractor (31) can proceed quickly (i.e., parameter update can be performed quickly and the loss convergence speed also increases as the number of parameters is reduced compared to the case where the parameters of the deep learning model (11) are updated all at once).

[0125] In the second learning stage, the detection system (10) can perform additional learning by connecting the original predictor (32) to the learned feature extractor (31). For example, the detection system (10) can configure an input sequence for a plurality of segments (e.g., 146, i.e., basic segments) and predict whether sleep apnea occurs in each of the plurality of segments (e.g., 146) through the feature extractor (31) and the predictor (32). In addition, the detection system (10) can update the parameters of the feature extractor (31) and the predictor (32) based on the loss (149) between the prediction result (147) and the correct label (148). Although Fig. 14 illustrates that the feature extractor (31) receives data extracted from one basic segment and one extended segment, as described above, it can be understood that the feature extractor (31) repeatedly receives data extracted from multiple basic segments and multiple extended segments to extract a feature sequence.

[0126] In some cases, the detection system (10) may proceed with the second learning step while freezing the feature extractor (31).

[0127] In addition, in some cases, the detection system (10) may extract first ECG signal samples having a low difficulty in detecting sleep apnea (e.g., ECG signal samples including a large number of segments in which complete apnea occurred) and second ECG signal samples having a high difficulty in detecting sleep apnea (e.g., ECG signal samples including a large number of segments in which shallow breathing occurred) from the ECG signal dataset and use them to perform the first and second learning steps. For example, the detection system (10) may perform the first learning step using the first ECG signal samples. Next, the detection system (10) may adjust the values ​​of the parameters of the learned feature extractor (31) based on the proportion (ratio) of the first ECG signal samples in the ECG signal dataset (i.e., the number of first ECG signal samples relative to the total number of samples) (e.g., multiplying the values ​​of all parameters by the proportion value). Alternatively, the detection system (10) may evaluate the performance (e.g., accuracy) of the feature extractor (31) to which the temporary predictor (141) is connected using the second ECG signal samples and adjust the values ​​of the parameters of the learned feature extractor (31) based on the performance evaluation results. Alternatively, the detection system (10) may adjust the values ​​of the parameters of the learned feature extractor (31) based on both the weight values ​​and the performance evaluation results (e.g., multiplying the values ​​of all parameters by the weight values ​​and the performance evaluation results). Next, the detection system (10) may proceed with a second learning step using the second ECG signal samples.

[0128] Hereinafter, a method for training a deep learning model according to some embodiments of the present disclosure has been described with reference to FIG. 14. As described above, a deep learning model (11) for detecting sleep apnea can be constructed by first training a feature extractor (31) by connecting a lightweight temporary predictor (141), and then performing additional training by connecting an original predictor (32) to the trained feature extractor (31). In this case, compared to training the deep learning model (11) all at once, parameters such as the feature extractor (31) and the predictor (32) can be updated quickly, so that the time cost and computing cost required for training the deep learning model (11) can be significantly reduced.

[0129] Below, the results of experiments conducted by the inventors of the present disclosure are briefly described.

[0130] In order to demonstrate the effectiveness of the technical idea of ​​the present disclosure, the inventors constructed a deep learning model (e.g., 11) for detecting sleep apnea using electrocardiogram signal samples (hereinafter, abbreviated as “samples”) of actual sleep apnea patients, and conducted an experiment to evaluate the performance of the deep learning model.

[0131] Specifically, the inventors implemented a deep learning model having the structure illustrated in FIGS. 5 to 8 and performed training and performance evaluation of the deep learning model using the 5-fold cross validation technique. The inventors performed training and performance evaluation of the deep learning model using a total of 1,465 samples, using the cross-entropy function as the loss function. Training of the deep learning model was performed using the method illustrated in FIG. 14.

[0132] The performance evaluation results are shown in Table 1 below. Those skilled in the art are likely already familiar with the concepts and measurement methods of the evaluation metrics listed in Table 1, so further explanation is omitted.

[0133] Metric Evaluation Results Accuracy 0.731±0.007 Specificity 0.620±0.025 Sensitivity 0.908±0.025 F1-score 0.731±0.006 AUC 0.892±0.010

[0134] Referring to Table 1, we can see that the deep learning model performs remarkably well regardless of the metric type. This demonstrates that using two features (i.e., feature 1 and feature 2) and an RNN-based predictor (see Figure 8), sleep apnea onset periods can be accurately detected from ECG signals.

[0135] The experimental results conducted by the inventors have been briefly described so far. Below, with reference to FIG. 15, an exemplary computing device (150) capable of implementing a detection system (10) according to some embodiments of the present disclosure will be described.

[0136] Figure 15 is an exemplary hardware configuration diagram showing a computing device (150).

[0137] As illustrated in FIG. 15, a computing device (150) may include one or more processors (151), a bus (153), a communication interface (154), a memory (152) for loading a computer program executed by the processor (151), and a storage (155) for storing a computer program (156). However, only components related to the embodiment of the present disclosure are illustrated in FIG. 15. Therefore, a person skilled in the art to which the present disclosure pertains will appreciate that other general components may be included in addition to the components illustrated in FIG. 15. That is, the computing device (150) may further include various components in addition to the components illustrated in FIG. 15. In addition, in some cases, the computing device (150) may be configured in a form in which some of the components illustrated in FIG. 15 are omitted. Hereinafter, each component of the computing device (150) will be described.

[0138] The processor (151) can control the overall operation of each component of the computing device (150). The processor (151) can be configured to include at least one of a CPU (Central Processing Unit), an MPU (Micro Processor Unit), an MCU (Micro Controller Unit), a GPU (Graphics Processing Unit), or any other type of processor well known in the art of the present disclosure. In addition, the processor (151) can perform operations for at least one application or program for executing operations / methods according to embodiments of the present disclosure. The computing device (150) can include one or more processors.

[0139] Next, the memory (152) can store various data, commands, and / or information. The memory (152) can load a computer program (156) from the storage (155) to execute operations / methods according to embodiments of the present disclosure. The memory (152) may be implemented as a volatile memory such as RAM, but the technical scope of the present disclosure is not limited thereto.

[0140] Next, the bus (153) can provide a communication function between components of the computing device (150). The bus (153) can be implemented as various types of buses such as an address bus, a data bus, and a control bus.

[0141] Next, the communication interface (154) can support wired and wireless Internet communication of the computing device (150). Furthermore, the communication interface (154) can also support various communication methods other than Internet communication. To this end, the communication interface (154) can be configured to include a communication module well known in the technical field of the present disclosure.

[0142] Next, the storage (155) can non-temporarily store one or more computer programs (156). The storage (155) can be configured to include non-volatile memory such as Read Only Memory (ROM), Erasable Programmable ROM (EPROM), Electrically Erasable Programmable ROM (EEPROM), flash memory, a hard disk, a removable disk, or any form of computer-readable recording medium well known in the art to which the present disclosure pertains.

[0143] Next, the computer program (156) may include one or more instructions that, when loaded into the memory (152), cause the processor (151) to perform operations / methods according to various embodiments of the present disclosure. That is, the processor (151) may perform operations / methods according to various embodiments of the present disclosure by executing the loaded one or more instructions.

[0144] For example, the computer program (156) may include instructions to perform an operation of acquiring an electrocardiogram signal, an operation of preprocessing a plurality of segments of the electrocardiogram signal to form an input sequence of a deep learning model (11), an operation of extracting a feature sequence from the input sequence through a feature extractor (31), and an operation of inputting the feature sequence to a predictor (32) to detect a segment in which sleep apnea has occurred among the plurality of segments.

[0145] As another example, the computer program (156) may include instructions to perform at least some of the steps / operations described with reference to FIGS. 1 through 14.

[0146] In the case illustrated, the detection system (10) according to some embodiments of the present disclosure may be implemented through a computing device (150).

[0147] So far, an exemplary computing device (150) capable of implementing a detection system (10) according to some embodiments of the present disclosure has been described with reference to FIG. 15.

[0148] The technical ideas of the present disclosure described with reference to FIGS. 1 to 15 so far can be implemented as computer-readable codes on a computer-readable recording medium. The computer-readable recording medium can be, for example, a removable recording medium (CD, DVD, Blu-ray disc, USB storage device, removable hard disk) or a fixed recording medium (ROM, RAM, computer-mountable hard disk). A computer program recorded on the computer-readable recording medium can be transmitted to another computing device via a network such as the Internet and installed on the computing device, thereby allowing the computer program to be used on the computing device.

[0149] Although all components constituting the embodiments of the present disclosure have been described as being combined or operating in combination as one, the technical concept of the present disclosure is not necessarily limited to such embodiments. That is, within the scope of the present disclosure, all of the components may be selectively combined and operated one or more times.

[0150] Although operations are depicted in the drawings in a particular order, this should not be understood to imply that the operations must be performed in the particular order depicted, or in any sequential order, or that all depicted operations must be performed to achieve the desired results. In certain circumstances, multitasking and parallel processing may be advantageous. Furthermore, the separation of the various components in the embodiments described above should not be understood to imply that such separation is absolutely necessary, and it should be understood that the program components and systems described may generally be integrated together into a single software product or packaged into multiple software products.

[0151] Although the embodiments of the present disclosure have been described with reference to the attached drawings, those skilled in the art will appreciate that the present disclosure can be implemented in other specific forms without changing the technical concepts or essential features thereof. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. The scope of protection of the present disclosure should be interpreted by the claims below, and all technical ideas within a scope equivalent thereto should be interpreted as being included within the scope of the technical ideas defined by the present disclosure.

Claims

1. A method performed by at least one computing device, A step of acquiring an electrocardiogram (ECG) signal; A step of preprocessing a plurality of segments of the electrocardiogram signal to form an input sequence of a deep learning model, wherein the deep learning model is formed by including a feature extractor and a predictor based on a recurrent neural network (RNN); A step of extracting a feature sequence from the input sequence through the feature extractor; and A step of inputting the feature sequence into the predictor to detect a segment in which sleep apnea occurs among the plurality of segments, How to detect sleep apnea.

2. In paragraph 1, The above feature extractor is implemented based on a CNN (Convolutional Neural Network) that performs a 1D convolution operation. How to detect sleep apnea.

3. In paragraph 1, The steps of configuring the above input sequence are: A step of extracting a plurality of RRI (RR Interval) and RPA (R-Peak Amplitude) values from a signal of a specific segment among the plurality of segments; and A step of generating data corresponding to the specific segment in the input sequence using the plurality of RRI and RPA values, How to detect sleep apnea.

4. In paragraph 1, The steps of configuring the above input sequence are: A step of preprocessing a signal of a specific segment among the plurality of segments to generate first data; A step of generating second data by preprocessing a signal of an extended segment corresponding to the specific segment, wherein the extended segment includes the specific segment and has a length longer than the specific segment; and A step of determining a pair of the first data and the second data as data corresponding to the specific segment in the input sequence, How to detect sleep apnea.

5. In paragraph 4, The above feature extractor comprises a first feature extractor, a second feature extractor, and a feature aggregator, The second feature extractor is configured at least in part differently from the first feature extractor, The step of extracting the above feature sequence is: A step of extracting a first feature from the first data through the first feature extractor; A step of extracting a second feature from the second data through the second feature extractor; and A step of generating an integrated feature by aggregating the first feature and the second feature through the feature aggregator, wherein the integrated feature is a feature corresponding to the specific segment in the feature sequence. How to detect sleep apnea.

6. In paragraph 5, The above first feature extractor is implemented with a deeper neural network than the above second feature extractor. How to detect sleep apnea.

7. In paragraph 5, The above first feature extractor is implemented based on ResNet (Residual neural network), The second feature extractor is implemented based on a CNN (Convolutional Neural Network) including an instance normalization layer. How to detect sleep apnea.

8. In paragraph 5, The steps for generating the above integrated feature are: A step of combining the first feature and the second feature; A step of analyzing the inter-channel correlation of the combined features to derive channel weights for the combined features; and A step of generating the integrated feature by reflecting the channel weights in the combined feature, How to detect sleep apnea.

9. In paragraph 8, The weight of a specific channel of the above combined feature is derived by analyzing the correlation between the specific channel and neighboring channels, The number of neighboring channels is determined based on the total number of channels of the combined feature. How to detect sleep apnea.

10. In paragraph 1, The above predictor is configured to include a first RNN block, a second RNN block, and a prediction layer, The above feature sequence includes a first feature corresponding to a first segment among the plurality of segments and a second feature corresponding to a second segment, The step of predicting the segment in which the above sleep apnea occurred is: A step of inputting the output value of the first RNN block for the first feature into the prediction layer to predict whether sleep apnea occurs in the first segment; and A step of inputting the output value of the second RNN block for the second feature into the prediction layer to predict whether sleep apnea occurs in the second segment, How to detect sleep apnea.

11. In paragraph 1, The learning process of the above deep learning model is: A step of training the feature extractor by connecting a temporary predictor to the feature extractor, wherein the temporary predictor is a neural network configured to receive a feature of a specific segment extracted from the feature extractor and predict whether sleep apnea occurs in the specific segment; and A step of performing additional learning by connecting the RNN-based predictor to the learned feature extractor, How to detect sleep apnea.

12. One or more processors; and A memory storing a computer program executed by one or more processors, The above computer program: The act of acquiring an electrocardiogram (ECG) signal; An operation of preprocessing a plurality of segments of the above electrocardiogram signal to form an input sequence of a deep learning model, wherein the deep learning model is configured to include a feature extractor and a predictor based on a recurrent neural network (RNN); An operation of extracting a feature sequence from the input sequence through the feature extractor; and Includes instructions for inputting the feature sequence to the predictor to detect a segment in which sleep apnea has occurred among the plurality of segments. Sleep apnea detection system.

13. Combined with a computing device, A step of acquiring an electrocardiogram (ECG) signal; A step of preprocessing a plurality of segments of the electrocardiogram signal to form an input sequence of a deep learning model, wherein the deep learning model is formed by including a feature extractor and a predictor based on a recurrent neural network (RNN); A step of extracting a feature sequence from the input sequence through the feature extractor; and In order to execute a step of inputting the above feature sequence into the above predictor and detecting a segment in which sleep apnea has occurred among the plurality of segments, stored in a computer-readable recording medium, Computer program.

Images