Method for classifying a polysomnography recording into defined sleep stages

Abstract

A method for classifying or categorizing a polysomnography recording into defined sleep tags.

Claims

1. A method for classifying a polysomnography recording into defined sleep stages, comprising the following steps: classifying the sleep of a human being into different sleep stages, wherein the sleep stages are identifiable by means of at least one datatype. of a first kind; collecting a plurality of information regarding bodily functions over a predetermined period of time in the form of data, with the data comprising at least one dataset of the datatype of the first kind; subdividing the collected data into time-dependent data blocks; manually selecting a limited number of training data blocks from the data blocks and assigning them to a sleep stage, wherein the training data blocks are selected in such a way that the data contained in the training block can each be uniquely assigned to a defined sleep stage; evaluating the dataset of the first kind of each training data block by means of a data preparation procedure; creating training objects, wherein each training object comprises the datasets of the last kind of a training data block evaluated by means of the data preparation procedure and the assignment of the training, data block to a sleep stage; transmitting the training objects to a support vector machine for creating a classification; transmitting at least some of the data blocks that were not selected as training data blocks to the support vector machine and automatically classifying said data blocks into the known sleep stages based on the data of the datatype of the first kind of the data blocks.

2. The method according to claim 1, characterized in that the dataset of the first kind comprises data of the following bodily functions: brain waves, cardiac activity, air flow of respiration, breathing sounds, in particular snoring sounds, eye movement patterns, electrical muscle activity in the chin area and on the lower leg.

3. The method according to claim 1, characterized in that at least one of the following measuring methods or measuring devices is used to collect the dataset of the first kind: electroencephalography, electrocardiography, microphone, air flow meter.

4. The method according to claim 1, characterized in that the dataset of the first kind comprises data of an electroencephalography.

5. The method according to claim 1, characterized in that the data preparation procedure comprises at least one of the following methods: cross-frequency coupling, entropy method, power spectral analysis and determination of heart rate variability when the dataset of the first kind comprises cardiac function data.

6. The method according to claim 1, characterized in that the cross-frequency coupling comprises a phase-amplitude coupling.

7. The method according to claim 1, characterized in that the collected data are divided into a predefined time interval, wherein in particular the time interval is in the range of 15 seconds to 5 minutes.

8. The method according to claim 1, characterized in that two to six, training data blocks are selected for each defined sleep stage.

9. The method according to claim 1, characterized in that the support vector machine comprises an algorithm that uses a non-linear basis kernel function.

10. The method according to claim 1, characterized in that the data on the bodily functions are collected in a sleep laboratory, wherein the data on the bodily functions are collected preferably during the second night in the sleep laboratory.

11. The method according to claim 1, characterized in that the data on the bodily functions are collected in a home environment.

12. The method according to claim 1, characterized in that the dataset of the datatype of the first kind consists of the data of an electroencephalography, and in that the evaluation of the dataset of the first kind of each training data block is performed by means of cross-frequency coupling with a phase-amplitude coupling.

13. The method according to claim 1, characterized in that the dataset of the datatype of the first kind consists of the data of an electroencephalography, and in that the evaluation of the dataset of the first kind of each training data block is performed by means of power spectral analysis.

14. The method according to claim 1, characterized in that the dataset of the datatype of the first kind consists of at least one of the following datatypes: data of an electroencephalography, respiratory flow, snoring sounds and in that the evaluation of the dataset of the first kind of each training data block is performed by means of an entropy method.

15. The method according to claim 1, characterized in that the dataset of the datatype of the first kind consists of the data of an electrocardiography and in that the data preparation procedure comprises a procedure to determine the heart rate variability.

16. The method according to claim 4, therein the dataset of the first kind comprises C3/C4 data of an electroencephalography.

17. The method according to claim 7, wherein the time interval is 30 seconds.

18. The method according to claim 8, characterized in that four training data blocks are selected for each defined sleep stage.

19. The method according to claim 12, wherein the data of an electroencephalography is C3/C4 data.

20. The method according to claim 13, wherein the data of an electroencephalography is C3/C4 data.

21. The method according to claim 14, wherein the data of an electroencephalography is C3/C4 data.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Preferred embodiments are explained in more detail with reference to the accompanying drawings, in which:

[0032] FIG. 1 shows a schematic representation of the flow of a method for classifying a polysomnography recording into defined sleep stages using electroencephalography (EEG) data in conjunction with a frequency coupling method;

[0033] FIG. 2 shows an overview of the usable data of the first kind and corresponding data processing methods;

[0034] FIG. 3 shows a schematic representation of the flow of a method for classifying a polysomnography recording into defined sleep stages using electroencephalography (EEG) data in conjunction with power spectral analysis;

[0035] FIG. 4 shows a schematic representation of the flow of a method for classifying a polysomnography recording into defined sleep stages using electroencephalography (EEG) data in conjunction with an entropy method;

[0036] FIG. 5 shows a schematic representation of the flow of a method for classifying a polysomnography recording into defined sleep stages using electrocardiography (ECG) data in conjunction with the determination of heart rate variability.

BEST WAY TO CARRY OUT THE INVENTION

[0037] FIG. 1 shows a schematic representation of the flow of a method for classifying a polysomnography recording into defined sleep stages based on electroencephalography (EEG) data in conjunction with a frequency coupling method.

[0038] In the first step of the method shown in FIG. 1, a person's sleep is divided into different sleep stages. Usually, sleep is divided into the five known stages, namely stage N1, stage N2, stage N3, REM stage and awake stage.

[0039] Each of these known stages can be identified on the basis of at least one datatype of the first kind. In the specific case, it is intended to automatically identify and classify the individual stages on the basis of the brain waves recorded by means of electroencephalography.

[0040] The next step is collecting a plurality of information regarding bodily functions during a person's sleep in the form of a well-known polysomnography recording in a sleep laboratory. Typically, a polysomnography recording lasts seven to eight hours.

[0041] The collected data is divided into time-dependent data blocks with a duration of 30 seconds. This can be done manually, i.e. by a person, or automatically by a computer or the like.

[0042] From said data blocks, a trained person or a specialist selects a limited number of training data blocks and assigns each of these selected training data blocks to a sleep stage, wherein the trained person or the specialist selects the training data blocks in such a way that the data contained in the training block can each be uniquely assigned to a defined sleep stage. Ideally, the trained person or specialist selects the same number of training data blocks for each sleep stage. It has been shown that the selection of four training data blocks per sleep stage is sufficient. However, it goes without saying that more or fewer training data blocks can be selected within the scope of the described method.

[0043] The polysomnography recording and thus the data blocks contain, among other things, the brain waves recorded by means of electroencephalography. The brain waves were recorded at different locations in the brain. For the further procedure of classifying a polysomnography recording into sleep stages, the data recorded at positions C3 and C4 on the head of a patient by means of electroencephalography are used (see illustration 1 in FIG. 1). Positions C3, C4 are those positions commonly referred to as C3, C4 in electroencephalography.

[0044] The data of each training data block obtained at the C3/C4 positions of an electroencephalography are analyzed using a data preparation procedure.

[0045] It is known that the frequency and amplitude of brain waves change during the different sleep stages. Each sleep stage is characterized by the presence respectively intensity or amplitude of different known frequency groups. Thus, the data displayed by the electroencephalogram at one position of the brain represent a superposition of different signals emitted by the brain in the form of brain waves. A simple frequency analysis of the collected data, for example in the form of a fast Fourier transform, due to the superimposed signals does not provide frequency sequences that can be clearly assigned to a steep stage.

[0046] For this reason, the data obtained at the C3/C4 positions of the electroencephalography are processed using cross-frequency coupling (see illustration 2 in FIG. 1). Surprisingly, it has been found that a cross-frequency coupling method with a phase-to-amplitude method is particularly suitable for assigning sleep stages to the data of an electroencephalogram.

[0047] From the data collected in the course of electroencephalography, two frequency groups are identified at the C3/C4 positions, the course and intensity of which can be described precisely by means of phase-to-amplitude coupling. By means of phase-to-amplitude coupling, the dependence between the amplitude of a higher-frequency signal and the phase of a lower-frequency signal is represented. The characteristic course of the frequency groups processed by means of phase-to-amplitude coupling can be clearly assigned to a seep stage.

[0048] The data of a data block obtained by means of cross-frequency coupling, in particular by means of phase-to-amplitude coupling, are correlated with the sleep stage determined by a skilled person and thus form a training object.

[0049] The training objects obtained from the selected data blocks are transmitted to a support vector machine to create a classification in the support vector machine (see Illustration 3 in FIG. 1).

[0050] An algorithm included in the support vector machine marks each data element as a point in n-dimensional space, where n represents the number of features. The algorithm has to calculate the best mean value between different separating straight lines in order to find the best common separating plane for a points, in this case a line with the maximum possible distance to all data points. The classification is performed by determining the so-called optimal hyperplane. As a next step, the algorithm looks for the hyperplane on which those data points with the smallest distance to said optimal hyperplane are located, the so-called support vectors. This distance is given the name Margin. The optimal separating hyperplane now maximizes the Margin to obtain clearly separated classification groups. The support vector machine thus divides the training data blocks into the specified sleep stages.

[0051] Then, the remaining data blocks that were not selected as training data blocks are transmitted to the support vector machine and an automatic classification of these data blocks into the known sleep stages based on the C3/C4 data of an electroencephalography is performed.

[0052] In a test phase, the described method was able to correctly assign the data blocks to sleep stages and thus achieve a hit rate of more than 93% (see Illustration 4 of FIG. 1).

[0053] A particularly accurate classification of data blocks which are not selected as training data blocks is achieved by using a non-linear basis kernel function in the support vector machine algorithm.

[0054] FIG. 2 shows an overview of the first kind of data that can be used in the present method and that are suitable for performing a classification, and corresponding data preparation procedures for evaluating, the first kind of data.

[0055] Instead of the data of an electroencephalography as used in connection with FIG. 1, the data of the respiratory flew, the data of the snoring and the data of an electrocardiography are also suitable as data of the first kind for carrying out the described method if suitable data preparation procedures are used.

[0056] Thus, as an alternative to the cross-frequency method used in connection with FIG. 1, the data of the first kind can also be evaluated by means of a power spectral analysis or an entropy method.

[0057] In power spectral analysis, the frequency-related power of a signal in a frequency band is specified. Power spectral analysis is suitable, for example, for data from an electroencephalography (see FIG. 3). In particular, the multi-taper method is suitable for this purpose.

[0058] As already mentioned above, classical frequency transformations such as the Fourier transform can only be applied to electroencephalography data insufficiently or not at all For example, a Fourier transform lacks the time-related reference to the respective frequencies. The multi-taper method generates such a time-frequency representation by multiplication in the frequency domain.

[0059] The entropy method is a non-linear dynamic analysis. The main principle of entropy methods is the quantification of information of a signal and of the probability of occurrence of certain patterns within a finite number of patterns and within a time series of the signal. The more information conveyed within a signal, the higher the entropy of the signal. While there are several kinds of entropy methods, in the context of sleep stage classification, the sample entropy method is particularly suitable, which is a modification of the approximate entropy method.

[0060] In the approximate entropy method, time series are examined regarding similar epochs, with more frequent and more similar epochs leading to lower values of approximate entropy. Thus, lower values of approximate entropy signify a high level of regularity of the signal and, conversely, high values of approximate entropy signify an irregular signal.

[0061] However, the approximate entropy method is dependent on the dataset length. Thus, in order to avoid the results being dependent on dataset length, an entropy method is used in which sequences that agree with themselves are not counted and which functions independently of dataset length. Said entropy method is the sample entropy method mentioned above, which is a modification of the approximate entropy method. The sample entropy method also has the advantage of being faster to perform.

[0062] It is particularly advantageous to use the sample entropy method in conjunction with electroencephalography data, as shown in FIG. 4. Respiratory flow and snoring data can also be used for this method.

[0063] In the event that the data is taken from an electrocardiography, a data preparation procedure that determines heart rate variability from the collected data is also suitable (see FIG. 5).

[0064] The interval between two heartbeats is usually defined as the time between the onsets of two contractions of the heart chambers. This onset of ventricular contraction is shown in the electrocardiogram as an R-wave, and the interval between two R-waves is called the RR-interval. The RR-intervals are usually not of equal length, but are subject to fluctuations. The quantification of these fluctuations is called heart rate variability (HRV).

[0065] The heart rate variability of a selected training data block determined from the data of an electrocardiography, together with the assignment to the sleep stage, already serves as a training object which can be transmitted to the support vector machine. The data blocks which are not selected as training data blocks can be classified into sleep stages by the support vector machine on the basis of the heart rate variability.

[0066] Even though in the described method the best hit rate respectively the best assignment of data blocks to sleep stages was achieved with the C3/C4 data using a cross-frequency method, a satisfactory hit rate was also achieved with the procedures using the entropy method or power spectral analysis. Except for the procedure using the snoring sounds as the first kind of data, the hit rates ware generally above 50%, in some cases well above 50%.

[0067] Heart rate variability is also suitable for the classification of sleep stages by means of the described method, Here, too, the hit rates are above 50%.