Self-supervised learning-based sleep stage classification model learning method and system using a small number of labels

Abstract

Provided are a method and a system for training a sleep stage classification model of a self-supervised learning infrastructure using a small number of labels. A sleep stage classification method executed by a computer system according to an embodiment includes inputting sleep test data into a sleep stage classification model of a self-supervised learning infrastructure, and classifying a sleep stage from the sleep test data using the sleep stage classification model of the self-supervised learning infrastructure. The sleep classification model of the self-supervised learning infrastructure may be one in which a pattern for sleep stage classification is learned from new sleep data by transfer learning by finely adjusting a weight value based on a representation learning model in which a representation is learned using sleep signal data.

Description

Technical Field

[0001] The following description relates to a technique for classifying sleep stages.

Background Art

[0002] Sleep occupies a very large proportion in daily life and has an important impact on the body. Since sleep disorders and sleep deprivation cause problems in daily life such as decreased concentration, research on sleep has been actively conducted. Sleep stage classification is an essential element in sleep research, and it has become possible to predict efficiently and accurately by using artificial intelligence. Sleep stage classification based on deep learning mainly used a method based on a supervised learning model, but recently, there has been an increasing interest in models using self-supervised learning.

[0003] In the sleep stage classification method using a supervised learning model, rules are utilized by using the given labels. Such a sleep stage classification method using a supervised learning model has several limitations. In order to improve the performance of the supervised learning model, it is realistically limited to produce labels every time for each dataset and proceed with learning, and there is a problem that the quality of rule learning also depends on the quality of the labels. In addition, the performance of the sleep stage classification method using a supervised learning model is limited to the dataset on which learning has been performed, so it is not appropriate.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Provided are a method and a system for classifying sleep stages with few labels by extracting a compressed representation that appropriately represents data from the signal itself by a self-supervised learning technique.

Means for Solving the Problems

[0005] A sleep stage classification method executed by a computer system includes a step of inputting sleep test data into a sleep stage classification model of a self-supervised learning base, and a step of classifying sleep stages from the sleep test data using the sleep stage classification model of the self-supervised learning base. The sleep classification model of the self-supervised learning base may be one in which a pattern for sleep stage classification is learned from new sleep data by transfer learning by finely adjusting a weight value based on a representation learning model in which a representation is learned using sleep signal data.

[0006] The sleep classification model of the self-supervised learning base may be composed of an adversarial generation model of an unsupervised learning base in order to minimize the influence on the class imbalance inherent in the sleep signal data.

[0007] The adversarial generation neural network model of the unsupervised base may receive inputs of electroencephalogram data and electrooculogram data in a preset time unit extracted from the sleep signal data, and be learned in a direction to maximize the mutual information amount between the signal generated from the generator and the class code.

[0008] The generator may receive a code for distinguishing classes and noise as inputs and generate class-specific data.

[0009] The sleep classification model of the self-supervised learning base may use an attention-based signal enhancement technique to generate a plurality of positive pairs for the sleep signal data, and perform pair-wise representation learning so that the distances between the generated plurality of positive pairs approach each other.

[0010] The sleep classification model of the self-supervised learning base may obtain an attention score for the sleep signal data, cover the obtained attention score with a random ratio of masks, and randomly drop the result of the attention score matrix, thereby obtaining a plurality of positive pairs covered with different masks.

[0011] Pair-wise representation learning may be performed by a method of reducing the negative cosine similarity between the generated positive pairs.

[0012] Pairwise representation learning may be performed by a method of constructing positive pairs for the overall sleep signal data to reduce the negative cosine similarity, or a method of constructing positive pairs for the sleep signal data in a preset signal unit to reduce the average of the negative cosine similarity.

[0013] A method for classifying a physical state executed by a computer system includes a step of inputting biological signal data into a physical state classification model of a self-supervised learning base, and a step of classifying a physical state from the biological signal data using the physical state classification model of the self-supervised learning base. The physical state classification model of the self-supervised learning base may be one in which a pattern for classifying a physical state stage is learned from new biological data by transfer learning by finely adjusting a weight value based on a representation learning model in which a representation is learned using biological data.

[0014] A computer system for sleep stage classification includes a data input unit that inputs sleep test data into a sleep stage classification model of a self-supervised learning base, and a sleep stage classification unit that classifies a sleep stage from the sleep test data using the sleep stage classification model of the self-supervised learning base. The sleep classification model of the self-supervised learning base may be one in which a pattern for sleep stage classification is learned from new sleep data by transfer learning by finely adjusting a weight value based on a representation learning model in which a representation is learned using sleep signal data.

[0015] The sleep classification model of the self-supervised learning base is composed of an adversarial generation model of an unsupervised learning base in order to minimize the influence on the class imbalance inherent in the sleep signal data. The adversarial generation neural network model of the unsupervised base may receive inputs of electroencephalogram data and electrooculogram data in a preset time unit extracted from the sleep signal data and be trained in a direction of maximizing the mutual information amount between the signal generated from the generator and the class code.

[0016] The self-supervised learning-based sleep classification model may generate a plurality of positive pairs for sleep signal data using an attention-based signal enhancement technique, and perform pair-wise representation learning so that the distances between the generated positive pairs approach each other.

[0017] The self-supervised learning-based sleep classification model may obtain an attention score for sleep signal data, and cover the obtained attention score with a random ratio of masks to randomly drop out the results of the attention score matrix, thereby obtaining a plurality of positive pairs covered with different masks from each other.

[0018] The pair-wise representation learning may be performed by a method of reducing the negative cosine similarity between the generated positive pairs.

[0019] The pair-wise representation learning may be performed by a method of constructing positive pairs for the overall sleep signal data to reduce the negative cosine similarity, and a method of constructing positive pairs for the sleep signal data of a preset signal unit to reduce the average of the negative cosine similarity.

Advantages of the Invention

[0020] By using the self-supervised learning-based sleep stage classification model, the sleep stage can be accurately classified even with a small number of labels.

Brief Description of the Drawings

[0021]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Embodiments for Carrying Out the Invention

[0022] Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

[0023] FIG. 1 is a diagram for explaining the operation of representation learning in one embodiment.

[0024] Sleep stage classification means classifying intervals defined by a sleep stage manual into specific stages based on polysomnography data. Polysomnography may include sleep examination data based on various electrical signals such as various electroencephalogram data, electrooculogram data, electrocardiogram data, electromyogram data, etc. during sleep. The computer system may automatically classify the sleep stage based on the sleep examination data. At this time, the computer system may be a sleep stage classification system that classifies the sleep stage. The computer system can predict the sleep stage with few labels by extracting a compressed representation that appropriately represents the data from the signal itself by a supervised learning technique.

[0025] The computer system may use an attention-based signal enhancement and a supervised learning model to learn effective representations from electroencephalogram data (EEG) and electrooculogram data (EOG) data among the sleep examination data, and classify the sleep stage using a small number of labels. The computer system may execute a representation learning stage of learning the overall characteristics of the sleep examination data using a representation learning model and a transfer learning stage of classifying the sleep stage with a small number of labels.

[0026] The representation learning model may be composed of a self-supervised learning framework based on a generative model. The representation learning model may extract electroencephalogram data and electrooculogram data in units of a preset time (for example, 30 seconds (1 epoch)) from the entire-time sleep examination data and use them as input data. In representation learning, in order to minimize the influence on the class imbalance inherent in the sleep examination data, an adversarial generative model based on an unsupervised learning framework is used as the basic structure.

[0027] First, the generator may receive the code c for distinguishing classes and the noise z as input data and generate class-specific data. At this time, in order to learn the latent factors for distinguishing classes, the adversarial generative model is learned in the direction of maximizing the mutual information between the signal generated by the generator and the class code. When learning the adversarial generative model, the objective function for maximizing the mutual information is as follows.

Equation

[0028] Next, for learning a unique representation for each class, pairwise representation learning based on a self-supervised learning framework may be performed. Pairwise representation learning may generate two positive pairs for the original signal data (electroencephalogram data, electrooculogram data), and the learning may proceed so that the distance between the positive pairs approaches. In order to generate two positive pairs for the original sleep signal data, a signal enhancement technique based on an attention framework may be used. First, when encoding the input signal, the attention score may be derived by the following method.

Equation

Number

[0029] Algorithm 1: Positive Pair Representation Learning Algorithm JPEG2025517151000005.jpg105115

[0030] As can be seen from Algorithm 1, two positive pairs can be generated by randomly masking the attention score matrix, and the learning can be advanced by reducing the negative cosine similarity between the positive pairs. The loss function for pair-wise representation learning that reduces the negative cosine similarity is as follows.

Number

[0031] After the two positive pairs are encoded, p 1 , p 2 are obtained by passing through the latent prediction model, and the obtained p 1 , p 2 are passed through the latent factor model Q to obtain q 1 , q 2 . sg is the "stop gradient" operation, which blocks the update of the model's learning parameters. As a result, when learning the latent representation that makes the two positive pairs similar, the model's learning parameters are updated only for the p 1 part, and q2 is only used to obtain the cosine similarity.

[0032] In addition, since the pattern of significant fluctuations seen within the signal of one epoch varies depending on the type of sleep stage, learning may be advanced by constructing positive pairs for the global loss function and the local loss function, respectively. In the case of the global loss function, learning may be advanced by constructing positive pairs for the overall part of the signal and reducing the negative cosine similarity. The global loss function is as follows.

Equation

[0033] In the case of the local loss function, the signal may be cut out in units of a preset time (e.g., 5 seconds), and learning may be advanced by constructing each positive pair and reducing the average of the negative cosine similarities. The local loss function is as follows.

Equation

[0034] The objective function of pair-wise representation learning including both the local loss function and the global loss function is as follows.

Equation

Equation

[0035] After representation learning, transfer learning may be performed for sleep stage classification. The transfer learning may be advanced by attaching one fully connected layer to the end of the latent factor model Q. The sleep stages may be classified by a method of fine-tuning using the weighted values of the model learned in the representation learning and the labels restricted to the weighted values of one fully connected layer.

[0036] Although the operations of representation learning and transfer learning for sleep stage classification have been described in the embodiments, the present invention is not limited thereto, and in addition to sleep stage classification, it can be similarly applied to body state classification.

[0037] FIG. 2 is a diagram showing an example of data utilization when performing representation learning and transfer learning in one embodiment.

[0038] The computer system may perform representation learning and transfer learning based on a self-supervised learning model. In the representation learning stage, without using labels, the computer system learns a representation that compresses data by performing unsupervised learning. In the representation learning stage, Sleep EDF, which is most frequently used among sleep data, may be used. In the transfer learning stage, the computer system fine-tunes the weighted values based on the representation learning model learned in the representation learning stage, and learns a sleep stage classification pattern using a small number of labels. In the transfer learning stage, learning is performed using new data that was not used in the representation learning, and thus it can be transferred and used for any data.

[0039] To evaluate the accuracy of sleep stage classification according to the embodiments, Sleep EDFx data and ISRUC data may be utilized. In the case of Sleep EDFx data, the sleep stages are classified using 10% and 20% of the data and their labels that were not used for feature learning of the sleep data. In the case of ISRUC data, the sleep stages are classified in a form of fine-tuning the weighted values using 10% and 20% of the labels of ISRUC for the model learned with 20% of Sleep EDFx data. To compare the performance, average accuracy and F1-macro may be used.

[0040] The following table shows the performance of the models for each data. To compare the performance of the sleep stage classification according to the embodiment with existing methods, a self-supervised learning-based model using a small number of labels and a supervised learning model using all labels were compared.

[0041] Performance comparison with self-supervised learning-based model [Table 1]

[0042] Performance comparison with supervised learning-based model [Table 2]

[0043] As can be seen from the experimental results, it can be confirmed that the results of the experiment according to the embodiment show better performance than existing methods even with only a small number of labels. In particular, it can be confirmed that when using 10% labels, it exhibits overwhelmingly better performance than other self-supervised learning-based models. Also, it can be confirmed that for the ISRUC data not used during representation learning, it shows better performance compared to existing self-supervised learning-based models and supervised learning-based models. Thereby, it can be understood that in terms of generalization performance, it shows consistent performance compared to existing learning models, and it can be confirmed that it is a model that can perform relatively accurate sleep stage classification even in a limited label situation.

[0044] FIG. 3 is a block diagram for explaining the configuration of a computer system in one embodiment, and FIG. 4 is a flowchart for explaining a method for classifying sleep stages in one embodiment.

[0045] The processor of the computer system 100 may include a data input unit 310 and a sleep stage classification unit 320. Components of such a processor may be representations of different functions that are executed by the processor according to control instructions provided by program code recorded in the computer system. The processor and components of the processor may control the computer system to execute steps 410 to 420 included in the sleep stage classification method of FIG. 4. At this time, the processor and components of the processor may be realized to execute instructions by the code of the operating system included in the memory and the code of at least one program.

[0046] The processor may load the program code recorded in the file of the program for the sleep stage classification method into the memory. For example, when a program is executed in the computer system, the processor may control the computer system to load the program code from the program file into the memory according to the control of the operating system. At this time, each of the data input unit 310 and the sleep stage classification unit 320 may be different functional representations of the processor for executing corresponding partial instructions among the program code loaded into the memory to execute steps 410 to 420.

[0047] In step 410, the data input unit 310 may input the sleep test data into the sleep stage classification model of the self-supervised learning base. At this time, different electrical signal data among the sleep (multivariate) test data may be input into the sleep stage classification model. For example, among the sleep test data, electroencephalogram data and electrooculogram data may be input into the sleep stage classification model.

[0048] In stage 420, the sleep stage classification unit 320 may classify the sleep stage from the sleep study data by using the sleep stage classification model of the supervised learning base. For example, the sleep stage classification unit 320 may classify the sleep stage into sleep stages including stage 1, stage 2, stage 3, stage 4, and stage 5. Alternatively, the sleep stage classification unit 320 may classify the sleep stage into a light sleep stage and a deep sleep stage.

[0049] FIG. 5 is a flowchart for explaining a method of classifying a physical state in one embodiment.

[0050] In stage 510, the computer system may input the biosignal data into the physical state stage classification model of the supervised learning base. At this time, the biosignal data may be health diagnosis result data or data measured in the user's daily life.

[0051] In stage 520, the computer system may classify the physical state from the biosignal data by using the physical state stage classification model of the supervised learning base. As an example, the physical state may be a state in which state information is classified according to the degree to which the user's body can move. The computer system may classify the health state of the user's body based on the user's health diagnosis result data such as the heart, brain (stroke, cerebral hemorrhage), lungs, kidneys, etc. Alternatively, the computer system may classify the physical health state of the user based on the movement data (of joints) measured in the user's daily life. At this time, the computer system may classify the physical state based on the movement of the body that can be performed according to the user's age.

[0052] The apparatus described above may be implemented by hardware components, software components, and / or combinations of hardware components and software components. For example, the apparatus and components described in the embodiments may be implemented using one or more general-purpose computers or special-purpose computers, such as a processor, a controller, an ALU (arithmetic logic unit), a digital signal processor, a microcomputer, an FPGA (field programmable gate array), a PLU (programmable logic unit), a microprocessor, or various devices capable of executing instructions and responding. The processing device may execute an operating system (OS) and one or more software applications executed on the OS. Further, the processing device may access data, record, manipulate, process, and generate data in response to the execution of the software. For the sake of convenience of understanding, it may be described as if one processing device is used, but those skilled in the art will understand that the processing device may include a plurality of processing elements and / or multiple types of processing elements. For example, the processing device may include a plurality of processors or one processor and one controller. Also, other processing configurations, such as a parallel processor, are possible

[0053] Software may include a computer program, code, instruction, or a combination of one or more of these, and may configure a processing device to operate as desired or instruct the processing device independently or collectively. Software and / or data may be embodied in any kind of machine, component, physical device, virtual equipment, computer recording medium or device for interpretation based on the processing device or for providing instructions or data to the processing device. Software may be distributed on a computer system connected by a network and recorded or executed in a distributed state. Software and data may be recorded on one or more computer-readable recording media.

[0054] The method according to the embodiment may be realized in the form of program instructions executable by various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed for the embodiment or may be usable and known to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy (registered trademark) disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specially configured to record and execute program instructions such as ROM, RAM, flash memory, etc. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes executable by a computer using an interpreter or the like.

[0055] As described above, the embodiments have been described based on limited embodiments and drawings. However, those skilled in the art will be able to make various modifications and variations from the above description. For example, even if the described technology is executed in an order different from the described method, and / or the components such as the described system, structure, device, circuit, etc. are combined or combined in a form different from the described method, and are opposed or replaced by other components or equivalents, appropriate results can be achieved.

[0056] Therefore, even if they are different embodiments, as long as they are equivalent to the claims, they belong to the appended claims.

Claims

1. A method for classifying sleep stages executed by a computer system, comprising: inputting sleep test data into a sleep stage classification model of a self-supervised learning foundation; classifying sleep stages from the sleep test data by using the sleep stage classification model of the self-supervised learning foundation ; the sleep classification model of the self-supervised learning foundation is one in which a weight value is finely adjusted based on a representation learning model whose representation is learned by using sleep signal data, and a pattern for sleep stage classification is learned from new sleep data by transfer learning . A method for classifying sleep stages, characterized by the above.

2. The sleep classification model of the self-supervised learning foundation is configured by an adversarial generation model of an unsupervised learning foundation in order to minimize the influence on the class imbalance inherent in the sleep signal data . The method for classifying sleep stages according to claim 1, characterized by the above.

3. The adversarial neural network model of the unsupervised foundation is configured to receive inputs of electroencephalogram data and electrooculogram data in a preset time unit extracted from the sleep signal data, and to be learned in a direction of maximizing the mutual information amount between the signal generated from a generator and a class code . The method for classifying sleep stages according to claim 2, characterized by the above.

4. The generator is configured to receive a code for distinguishing classes and noise as inputs and generate class-specific data . The method for classifying sleep stages according to claim 3, characterized by the above.

5. The sleep classification model of the self-supervised learning foundation is configured to use an attention-based signal enhancement technique to generate a plurality of positive pairs for the sleep signal data, and perform pair-wise representation learning so that the distance between the generated plurality of positive pairs approaches . The method for classifying sleep stages according to claim 1, characterized by the above.

6. The sleep classification model of the self-supervised learning foundation is configured to obtain an attention score for the sleep signal data, cover the obtained attention score with a random ratio of masks, and randomly drop the result of the attention score matrix, thereby obtaining a plurality of positive pairs covered with different masks . The method for classifying sleep stages according to claim 5, characterized by the above.

7. The pair-wise representation learning is performed by a method of reducing the negative cosine similarity between the generated positive pairs . The method for classifying sleep stages according to claim 5, characterized by the above.

8. The pair-wise representation learning is performed by a method of constructing positive pairs for the overall sleep signal data to reduce the negative cosine similarity and a method of constructing positive pairs for the sleep signal data of a preset signal unit to reduce the average of the negative cosine similarity. The sleep stage classification method according to claim 5, characterized by the above.

9. A body state classification method executed by a computer system, comprising: inputting the biological signal data into a body state stage classification model of a self-supervised learning base; and classifying the body state from the biological signal data by using the body state stage classification model of the self-supervised learning base. The body state stage method is characterized by including: The body state stage classification model of the self-supervised learning base is obtained by finely adjusting the weighted values based on a representation learning model in which the representation is learned using biological data, and learning a pattern for classifying the body state stage from new biological data by transfer learning. The body state stage method is characterized by the above.

10. A computer system for sleep stage classification, comprising: a data input unit that inputs sleep examination data into a sleep stage classification model of a self-supervised learning base; and a sleep stage classification unit that classifies the sleep stage from the sleep examination data by using the sleep stage classification model of the self-supervised learning base. The computer system is characterized by including: The sleep classification model of the self-supervised learning base is obtained by finely adjusting the weighted values based on a representation learning model in which the representation is learned using sleep signal data, and learning a pattern for classifying the sleep stage from new sleep data by transfer learning. The computer system is characterized by the above.

11. The sleep classification model of the self-supervised learning base is configured by an adversarial generation model of an unsupervised learning base in order to minimize the influence on the class imbalance inherent in the sleep signal data. The adversarial generation neural network model of the unsupervised base is trained in a direction of maximizing the mutual information amount between the signal generated from the generator and the class code by receiving the input of the electroencephalogram data and electrooculogram data of a preset time unit extracted from the sleep signal data. The computer system according to claim 10, characterized by the above.

12. The sleep classification model of the self-supervised learning base is Generating a plurality of positive pairs for the sleep signal data using an attention-based signal enhancement technique, and performing pair-wise representation learning so that the distance between the generated plurality of positive pairs approaches each other The computer system according to claim 11, characterized in that

13. The sleep classification model of the self-supervised learning base Obtaining an attention score for the sleep signal data, covering the obtained attention score with a random ratio of masks, and randomly dropping the results of the attention score matrix to obtain a plurality of positive pairs covered with different masks from each other The computer system according to claim 12, characterized in that

14. The pair-wise representation learning Is performed by reducing the negative cosine similarity between the generated positive pairs The computer system according to claim 12, characterized in that

15. The pair-wise representation learning Is performed by a method of constructing positive pairs for the overall sleep signal data to reduce the negative cosine similarity and a method of constructing positive pairs for the sleep signal data of a preset signal unit to reduce the average of the negative cosine similarity The computer system according to claim 12, characterized in that

Images