WO2020047569A1

WO2020047569A1 - Method for detecting breathing stops

Info

Publication number: WO2020047569A1
Application number: PCT/AT2019/060280
Authority: WO
Inventors: Carmina CORONEL; Christoph Wiesmeyr; Heinrich Garn; Bernhard Kohn; Alexander Schindler; Anahid NAGHIBZADEH-JALALI
Original assignee: Ait Austrian Institute Of Technology Gmbh
Priority date: 2018-09-05
Filing date: 2019-08-29
Publication date: 2020-03-12
Also published as: AT520925B1; AT520925A4

Abstract

The invention relates to a method for detecting breathing stops of a person, wherein: while the person sleeps, a breathing movement measurement value (M_B), which characterizes the movements of the person during breathing, and an oxygen saturation value (M_O2), which indicates the relative oxygen concentration in the blood of the person, are determined; the measurement values (M_B, M_O2) thus determined are correlated with one another, and a breathing stop is detected if the breathing movement measurement value (M_B) decreases during a first time period (T_A1, T_A2,...) and additionally the oxygen saturation value (M_O2) decreases during a second time period (T_B1, T_B2,...) following the first time period (T_A1, T_A2,...).

Description

Method of detection of breathing interruptions

The invention relates to a method for the detection of misfiring of a person. A large number of different possibilities for the detection of misfires are known from the prior art, which are based on different measuring principles. Typically, a polysomnography measurement is performed as part of an examination in the sleep laboratory. The air flow and chest movements are usually measured, among other things. In addition, oxygen saturation in a person's blood is usually also measured. If there is a misfire, the air flow and possibly also the intensity of the chest movements decrease. Subsequently, the oxygen saturation in the person's blood typically drops significantly over a period of 10 seconds to 1 minute. However, it is not yet known from the prior art to systematically record oxygen desaturation in sleep laboratory applications for the detection of complete interruptions in breathing (apneas) or systematically with other measurement principles, such as three-dimensional imaging methods of the patient, which do not hinder or disturb the patient during sleep to combine. If the person then begins to breathe normally again, the oxygen saturation in the person's blood increases again.

In addition, further procedures for determining respiratory movements are known from the prior art, which are based on distance measurements and in which the movements of the person's chest or abdomen are determined. If movements of the surface of the person or of the blanket that is located on the person are determined that take place with a predetermined frequency, this can be concluded from the person's breathing activity.

In addition, possibilities are also known for determining the breathing sounds of a person and, based on these breathing sounds, for concluding that the person is breathing.

A major problem in the prior art is that polysomnography is very cost-intensive and burdensome for patients and disturbs sleep. Since distance measurements as well as noise and oxygen desaturation can be measured with non-contact or minimally stressful sensors, a detection of misfires based on the fusion of these sensors is the Consider polysomnography in practical terms, in particular the patient is not disturbed in his sleep or only to a very small extent. In order to be able to implement robust and reliable detection of misfires, the methods must advantageously be coordinated with one another.

The invention is therefore based on the object of achieving an improved detection of misfires in a person which is more reliable and has a certain tolerance with regard to the individual, possibly faulty measurements. At the same time, the patient should not be disturbed while sleeping.

The invention solves this problem in a method for the detection of misfiring of a person with the characterizing features of claim 1. It is provided

- that the following measured values are continuously determined while the person is asleep:

a breath movement reading that characterizes the person's movements during breathing, and

an oxygen saturation value indicating the relative oxygen concentration in the person's blood, and

- That the measured values determined in this way are correlated with one another and a misfire is detected if the measured value of breathing movement drops during a first period and additionally the oxygen saturation value decreases during a second period that follows the first period.

The different effects of the person's breathing misfires on the individual measurement methods regarding the oxygen saturation value and the breath movement measured value are advantageously coordinated with one another, so that breath misfires are only detected if these are considered probable on the basis of the measured values obtained in both measurements. False alarms, which are only due to the unreliable functioning of one of the two measurement methods, are effectively prevented with this procedure. The patient is not hindered from sleep during these measurements. In particular, there is also the possibility of carrying out the measurement without the patient having to be wired for this, which also opens up the possibility of home applications or applications without supervision by specialist personnel and thus contributes significantly to reducing costs. A particularly advantageous procedure with which the time or period of the breathing misfire can be determined advantageously and reliably provides that the breathing mistake is considered to be within the first period. In addition, this also enables a temporal correlation of the breathing interruptions determined in this way with a large number of further measured values that may have been recorded during sleep, for example EEG measured values, video images, acoustic recordings, thermal images or similar recordings of the sleeping person. An advantageous procedure in which a large number of false positively detected misfires can be rejected provides that a misfire is only recognized if the second period of time lies within a certain period of time, in particular between 10 seconds and 60 seconds, after the first period . A further advantageous combination with a further measurement variable, which provides reliable information about breathing noises, provides that a sound measurement value is continuously determined during sleep of the person, which indicates the volume of the breathing noises emitted by the person, and a search is made for respiratory events that are characteristic of the occurrence of breathing noises, whereby it is examined whether the first period of time is free of breathing events and in this case an additional confirmation is made that breathing interruptions were present within the first period of time.

A numerically particularly stable procedure for the detection of breathing interruptions, which delivers reliable results, provides that the decrease in the measured value of the respiratory movement is recognized within a first period of time by

an envelope is determined for the temporal course of the measured breath movement value,

- A number of maxima of the envelope is determined, and

a time range between two maxima, in which the envelope falls below a predetermined threshold value, in particular determined by the values of the respective maxima, is determined as the first breath-free period.

A particularly robust detection of maxima in the envelope of the

Respiratory movement measurement value provides that a maximum is only used as such if

- Its time interval to the previous or subsequent maximum exceeds a predetermined time threshold and / or - His respiratory movement measured value exceeds a threshold value determined by the preceding or following maximum.

A numerically particularly efficient method for finding periods with low oxygen saturation provides that a second period of less

Oxygen saturation is determined by

Maxima are determined in the course of oxygen saturation,

- a search is made for a period of time when the oxygen saturation has dropped,

- in which the duration of the drop in oxygen saturation is at least a predetermined time threshold, and

at the end of which the oxygen saturation has decreased by at least 3% compared to the value of the maximum,

and such a time range is recognized as a second time period.

A particularly robust detection of maxima in the course of oxygen saturation provides that a maximum is only used as such if

- Its time interval to the previous or subsequent maximum exceeds a predetermined time threshold and / or

- Its saturation value exceeds a threshold value determined by the preceding or following maximum.

A further advantageous procedure for the detection of breathing interruptions can also be carried out by supplying the determined breath movement measured values and the determined oxygen concentration in the blood to a neural network. It is provided that for each of a number of time intervals, in particular of a duration between 0.5 and 5 seconds, a feature vector is determined from the measured values, that from the respiratory movement measured values recorded in the relevant time interval, a number of with regard to their position the feature vector entries specified in the feature vector are determined, and a number of feature vector entries specified with regard to their position in the feature vector is determined from the oxygen saturation values recorded in the relevant time window, that a time window comprising a predetermined number of features - Vectors or time intervals are specified such that the duration of the time window is selected such that it is preferably between 10 and 120 seconds so that the feature vectors V ₁ assigned to the time window _; V ₂ , ... are transferred to a previously trained neural network, which in particular has at least one recurrent Layer, preferably a recurrent LSTM layer, and that a temporal sequence of individual probability values is obtained by means of the neural network, which represents the probability of a misfire during the relevant time window or in relation to the relevant time interval for which a feature vector is valid is to specify.

A particularly advantageous procedure, with which errors are avoided at the edges of the window, provides that the time windows overlap one another at least partially, in particular by half, so that everyone is characterized by a feature vector or affected by it Two at a time

Probability values are determined which indicate the probability of a misfire at the relevant point in time, and that the two probability values obtained in this way for a feature vector or a point in time, in particular weighted, are averaged.

For this purpose, it can in particular be advantageously provided that the weighting of the separately determined and weighted added probability values is specified in such a way that values that are in the middle of the time window are weighted more strongly than values that are at the beginning of the time window.

In order to be able to localize breathing interruptions in time, it can be provided that time ranges in which the determined probability value for the presence of breathing interruptions exceeds a predetermined comparison threshold value, in particular determined by a moving average of the surrounding probability values, possibly multiplied by a predetermined factor are detected by misfires.

A particularly advantageous, numerically simple possibility for determining probability values provides that the neural network for determining the probability value has one or more further layers which are connected upstream or downstream of the LSTM layer and that the neural network ends in a single neuron, at the exit of which the probability value is obtained.

A program for carrying out a method according to the invention can be stored on a data carrier in a particularly advantageous manner. Two advantageous embodiments and further developments of the invention are shown in more detail with reference to the following drawing figures.

1 shows a diagram of measured values which are usually determined in the sleep laboratory. 2 shows the formation of an envelope curve for the temporal course of the measured breath movement value. 3 shows the determination of maxima of the envelope. FIG. 4 shows the determination of the first time periods that indicate the presence of breathing interruptions. 5 shows an oxygen saturation value and a smoothed time profile of the oxygen saturation value. 6 and 7 show the determination of maxima as well as of second time periods, which indicate the presence of breathing interruptions. 8 shows the joint consideration of first and second time periods for determining time periods in which breathing interruptions are present.

FIG. 9 schematically shows the procedure for forming time intervals in a second embodiment of the invention. 10 shows the formation of feature vectors for the individual time intervals and the formation of time windows. FIG. 11 shows a neural network with a recurrent LSTM layer for calculating a probability value for the existence of misfires. 12 shows the formation of cross-window probability profiles as well as the calculation of time ranges with misfires.

1 shows the individual signals determined in the course of the examination of the sleeping person according to the prior art. 1 shows function graphs arranged one above the other for the temporal course of the pressure F in the area of the nose, the temperature F _T in the area of the mouth and nose, measured breath movements, M _BI and M _B2 for measuring the breath movements via a chest or abdominal belt and the oxygen saturation M _Q2 in the subject's arterial blood. In the central region shown in FIG. 1, a breathing misfire can be seen in the individual measured value signals between the times A and A ₂ . Some of the measured values shown here provide significant information about breathing interruptions, but have the disadvantage that they are usually determined by cable, that is to say sensors are attached to the body of the person which significantly impair the person's sleep. Furthermore, a measured breath movement value is continuously determined. In contrast to the chest and abdominal belts in polysomnography, the breath movement measured value is specified via a distance measuring device, which measures the distance from the chest or abdomen area of the person concerned or a blanket lying above the chest and / or abdomen measurement area of the person. A "time of flight sensor" can typically be used as such a distance measuring device.

The oxygen saturation can be determined, for example, in the area of a finger. The oxygen saturation M _Q 2 is determined, for example, at intervals of 0.5 to 5 seconds. The oxygen saturation signal M _Q 2 is present as a temporally discrete signal, the sensor also determining discretized signals with respect to the oxygen saturation measured value M _Q 2. Subsequently, a first embodiment of the invention is presented, which manages with an oxygen saturation signal M ₀ 2 and a breath movement measurement value M _B , in order to be able to make reliable statements about breathing interruptions and their occurrence over time. In a first step, with reference to FIGS. 2 to 4, measures are described with which the presence of a breath misfire can be detected from a measured breath movement value. In this case, an envelope curve Fl is determined from the measured breath movement value, which is limited in terms of its oscillation (FIG. 2). For this purpose, a curve with limited oscillation enveloping the measured breath movement value from above and a curve with limited oscillation enveloping the measured breath movement value from below are calculated and the envelope curve F1 is formed as the difference between these two curves.

The time and value of the envelope curve Fl are determined in the individual maxima M1, M2, ... of the envelope curve Fl, which are shown in more detail in FIG. 3. A maximum is then considered to be detected and is only used as such for the further procedure if its time interval to the preceding or following maximum in each case exceeds a predetermined time threshold. This time threshold can be between 5 and 10 seconds. Furthermore, as an alternative or in addition, the maxima can also be limited to those signal values in which the envelope curve exceeds a threshold value determined by the preceding or following maximum. For example, can be set in this way that the value of a maximum may be at least 90% of the value of the immediately preceding maximum.

If a curve of the envelope curve H of the breath movement measured value is determined between two maxima, in which the value of the envelope curve H for a predetermined first time period TAI,. . . , T A6 - in the present exemplary embodiment of at least 10 seconds - has a value which is smaller than a proportion of - in the present exemplary embodiment 80% - the mean value of the two maxima, then the first time period between the two maxima in which this threshold value is undershot, determined as breathable based on the measured breath movement value. As in the previous exemplary embodiment, a period of between 5 and 30 s can be selected. A threshold of between 50% and 90% of the mean value of the function values of the envelope curve H in the two successive maxima M. Can be used as the threshold value, which must be undercut for the detection of a breathing misfire during the first time period T _Ai , ..., T A6 _{1 ;} M ₂ can be used.

FIG. 4 shows first time periods T _Ai , ..., T _A6 in the signal of the breath movement measured value M _B , in which breathing misfires are considered to be recognized. In a second step, the discretely determined oxygen saturation signal M ₀ 2 is converted piece by piece into a smooth signal, for example by piece by piece linear interpolation or by spline interpolation with square or cubic splines. The discretely determined signal and the smoothed and linearly interpolated signal are shown in more detail in FIG. 5.

The time and the oxygen saturation M _Q 2 in the individual maxima are then determined in the smoothed signal. Between two immediately consecutive maxima, a time range for the drop in oxygen saturation M _Q 2 is sought, in which the duration of the drop in oxygen saturation M _{02 is} at least a predetermined second time period T _Bi , ·, T _B4 - in this

Embodiment 10 seconds - and in which the oxygen saturation M _Q 2 compared to the value of the maximum by a predetermined value - of 3% in this embodiment - has dropped. Instead of the parameters of the method specifically used here, other different threshold values can also be selected. The second time period T _B1 ,... T _B4 , for which the threshold value has to be undershot so that a breathing misfire can be detected, can be between 5 and 30 seconds. The drop in oxygen saturation M _Q 2 that will be present must, so that a breathing mistake is considered detected, can be between 2% and 10% relative oxygen saturation M ₀₂ .

6 there are two maxima T _1; T ₃ and the minumum T ₂ between them are shown. In the present case, the oxygen saturation at the time J- _{\ has} the value 93%, the oxygen saturation at the time T _{3 has} the value 95%. At the time T _{2 of} the minimum, the oxygen saturation is 87%.

The saturation difference between the smaller of the two maxima and the minimum of the oxygen saturation in the present case is 93% -83% = 10%. The saturation difference determined in this way is compared with a threshold value which can be between 3% and 15%; in the present case, a threshold value of 5% is selected as an example, so that the saturation difference of 10% exceeds this threshold value. The time interval T ₃ -T between the two maxima is 20 seconds in the present case. For this time interval it is now examined whether it lies within an interval of between 5 seconds and 90 seconds. The lower limit of this interval can be between 2 seconds and 20 seconds. The upper limit of this interval can be between 30 seconds and 200 seconds.

Since, in the present case, on the one hand the saturation difference exceeds the predefined threshold value and on the other hand the time interval T3-T1 lies within the predefined interval, the period between the two maxima is evaluated as the second period T _Bi , ..., T _{B4 of} a breathing misfire.

FIG. 7 shows a number of second time periods T _Bi , ·, T _B4 , in which a reduced oxygen saturation M _Q 2, which indicates a misfire, is detected. After both the respiratory movement measured value and the oxygen saturation signal M _Q 2 have identified first and second time periods T _Ai,, T _B1 , ... in which the respective signal indicates the presence of a

Interrupting breaths, it is subsequently examined whether the first and second periods T _Ai , ..., T _B1 , ... determined in this way correlate with one another in time (FIG. 8). It searches for interrelated first and second time periods T _Ai , ..., T _B1 , ..., which are shifted from each other by a predetermined distance DT5, with a first time period T _{A3 of} a misfire in

Respiratory movement measured value M _{B is} determined and then after a second, occurring approximately a delay period of between 10 and 60 seconds later Time period T _{Bi is} sought in the oxygen saturation measurement _signal M ₀₂ , which also indicates the presence of a misfire. If such correlating first and second time periods T _Ai , T _{Bi can be found} in the time profiles of the oxygen saturation value M ₀₂ and the breath movement measured value, then a person's breathing misfire is detected during the first time period T _Ai . In this way, the periods T _Ai , T _Bi and the periods T _A3 , T _{B3 can} be assigned to each other. Since in the present case the second period T _B2 lies before the temporally close first period T _A2 , an assignment cannot be made. A preferred development of the invention provides that a sound measurement value M _akk is additionally determined while the person is _asleep , it being continuously examined whether the volume of the breathing noises emitted by the person exceeds a predetermined threshold value. If necessary, the sound measurement value M _akk can be subjected to filtering beforehand. You can also search for breathing events or breathing noises that are characteristic of the occurrence of breathing. If it can be determined that the first period T _A3 , in which the

Breathing movement measured value M _B falls below a predetermined level, and is also free of acoustically identified breathing events, a further confirmation is given that breath misfires exist within a first period of time.

In a second preferred embodiment of the invention, shown in FIGS. 9-12, as in the first embodiment shown above, a breath movement measurement value and an oxygen saturation value M ₀₂ are continuously determined to detect a person's breathing interruptions while the person is sleeping. The total duration of monitoring sleep is broken down into a number of

Time intervals l _1; l ₂ , ... divided, which in the present embodiment

Have interval duration of about one second (Fig. 9). In principle, different interval durations, for example between 0.5s and 5s, can also be selected. The time intervals l _1; l ₂ , ... are not overlapping in the present exemplary embodiment and are specified without any time intervals.

The individual signals are used for the individual time intervals l _1; l ₂ , ... feature values determined, the individual parameters of the signal curve in the respective time interval l _1; l ₂ , ... describe. It is entirely possible that values of the signals are used that are in close proximity to one another but outside the respective time interval h, l ₂ ,. As will be shown later, the individual feature values can be determined in different ways. The individual feature values determined in this way, which were each determined for the same time interval \ ^, l ₂ , ..., become feature vector entries for a feature vector V _1; V ₂ , ... summarized, which in turn the time interval l _1; l ₂ , ... is assigned (Fig. 10). The order of the individual feature vector entries in the feature vector V _1; V ₂ , ... is basically arbitrary, but is used when creating all feature vectors V _1; V ₂ , ... are retained for the individual time intervals \ l ₂ , ..., so that feature values which are in the individual time intervals l _1; l ₂ , ... were each determined in the same way, in each case at the same location in the feature vector V _1; V ₂ , ... stand.

The individual time intervals l _1; l ₂ , ... for time windows F _1; F ₂ , ... summarized, wherein a time window F _1; F ₂ , ... each comprises a predetermined number of time intervals. Every time window F _1; F ₂ , ... can therefore also be a number of feature vectors V _1; V ₂ , ... are assigned if they are assigned to a time interval \ l ₂ , ... which corresponds to the time window F _1; F ₂ , ... is assigned. Through the chronological sequence of the time intervals, l ₂ , ... in the respective time window F _1; F ₂ , ... there is also a sequence of the time intervals l _1; l ₂ , ... assigned feature vectors V _1; V ₂ , ....

The duration of the time window F _1; F ₂ , ... is typically between 10 and 120 seconds. The duration of the time window is preferably greater than the expected delay time, with which a reduction in the oxygen saturation value M ₀ 2 typically occurs when there is a breathing misfire.

The individual feature vectors V _1; V ₂ , ... are connected to a feature vector V _1; V ₂ , ... transferred trained neural network. This neural network (FIG. 11) contains at least one recurrent LSTM layer. The basic structure of such a layer of a neural network is described, for example, in Felix A. Gers, Jürgen Schmidhuber, Fred Cummins: Learning to Forget: Continual Prediction with LSTM In: Neural Computation (journal) vol. 12 issue 10, pp. 2451-2471, 2000. The input layer N, the neural

Network N points for each entry of feature vector V _1; V ₂ , ... each have an input neuron N _n , ..., N _i5 . The neural network N has an output or an output neuron N _out , to which a value is present that represents the probability P for the presence of a misfire in the presence of a feature vector V _{1 present} at the input N _; V ₂ , ... indicates. Alternatively, the neural network can also contain as many output neurons as there are time intervals in a time window. The neural network typically has N for determining the Probability value P one or more further layers N ₂ , which are upstream or downstream of the LSTM layer.

To determine the individual probability values P, the individual feature vectors V _1; V ₂ , ... of the respective window F _1; F ₂ , ... in the order given to them by the temporal sequence of their recording to the neural network N (Fig. 12). By transferring the feature vectors V _1; V ₂ , ... to the neural network N is not only for each time interval l _1; l ₂ , ... receive a probability value P for the presence of a misfire, but the neural network N also changes its internal state due to its LSTM layer, so that even events distributed over time can be easily recognized.

Instead of an LSTM layer in a neural network, basically other layers or layers can also be used. For example, it may be sufficient if, instead of an LSTM layer, a fully connected layer (fully connected layer) or a recurrent layer that is not of the LSTM type is used, such as a "gated recurrent unit".

After all feature vectors V _1; V ₂ , ... were passed in their order to the neural network N, a time output signal Pi, P ₂ , P ₃ is obtained at the output N _{out of} the neural network N, which represents the probability P of a misfire at the time interval h, l ₂ , ... of the respective feature vector V _1; V ₂ , ... or the time window F _1; F ₂ , ... indicates. Depending on the choice of the size of the time window F _1; F ₂ ,... Can in particular be provided that they overlap each other by a certain time span, so that each time interval l _1; l ₂ , ... always in at least two time windows F _1; F ₂ , ... lies. In this case, a feature vector V _1; V ₂ , ... characterized or affected by this point in time at least two probability values P _1; P ₂ , ...; P ₂ , P ₃ , ... are determined, the probability of a misfire to the concerned

Specify the time. The two so for a feature vector V _1; V ₂ , ... or probability values obtained at a point in time can be averaged to improve the probability estimate. The addition of the probability values P can preferably be weighted. The weighting of the separately determined probability values P can depend on the time position of the probability value P to be weighted in its Time window can be specified. So probability values P, which are in the middle of the time window F _1; F ₂ , ... are weighted more than values that are at the beginning or end of the time window F _1; F ₂ , ... are. In order to be able to effectively detect the presence of a misfire, the determined probability value P can be compared with a threshold value P _th . Since the probability value P does not necessarily indicate the actual probability in the present case, but may also be influenced by different parameters that are essentially the same for the sleep laboratory arrangement in question, a comparison _threshold value P _th for the probability value P cannot be specified in general. A moving averaging of the relevant probability value signal P enables an adjustment and determination of a comparison threshold value. The comparison _threshold value P _th can either be directly equated with the moving average of the probability value P or can be set to a value predetermined by the moving average, for example between 120% and 150% of the moving average. If the probability value signal exceeds the comparison _threshold value P _th , a misfire is detected and possibly assigned to the relevant time range B.

Time ranges B in which the ascertained probability value P for the presence of breathing interruptions exceeds a predetermined comparison _threshold value P _th , in particular determined by a moving average of the preceding probability values, are determined as time range B of breathing interruptions, as shown in FIG. 12.

Claims

Claims:

1. A method for detecting a person's breathing interruptions,

- The following measured values are continuously determined while the person is asleep:

- a breath movement reading (M _B ) characterizing the person's movements during breathing, and

- an oxygen saturation value (M ₀₂ ) indicating the relative oxygen concentration in the person's blood, and

- The measured values (M _B , M02) determined in this way are correlated with one another

Breathing interruptions are detected if the breath movement measured value (M _B ) decreases during a first period (T _Ai , T _A2 ,) and additionally during a second period (T _B1 , T _B2 , ...) that corresponds to the first period (T _Ai , T _A2 ,) follows, the oxygen saturation value (M ₀₂ ) drops.

2. The method according to claim 1, characterized in that the breathing interruptions are regarded as lying within the first period (T _Ai , T _A2 , ...).

3. The method according to any one of the preceding claims, characterized in that a misfire is detected only when the second period (T _B1 , T _B2 , ...) in a certain period, in particular between 10 seconds and 60 seconds after first period (T _A1 , T _A2 , ...).

4. The method of claim 1 or 2, wherein in addition, a sound measurement value (M _akk ) is continuously determined during sleep of the person, which indicates the volume of the breathing noises emitted by the person, and is searched for breathing events that are characteristic of the occurrence of breathing noises are,

it is examined whether the first time period (T _Ai , T _A2 , ...) is free of breathing events and in this case an additional confirmation is made that within the first time period (T _Ai , T _A2 , ...) breath failures were present.

5. The method according to any one of the preceding claims, characterized in that the decrease in the measured breath movement value (M _B ) within a first period (T _Ai , T _A2 , ...) is detected by

an envelope curve (H) is determined for the temporal course of the measured breath movement value (M _B ),

- A number of maxima (M _1; M ₂ , ...) of the envelope (H) is determined, and - A time range between two maxima (M _1; M ₂ , ...) in which the envelope falls below a predetermined threshold value, in particular determined by the values of the respective maxima (M _1; M ₂ , ...), as the first breathable period is determined.

6. The method according to claim 5, characterized in that a maximum is only used as such if

- its time interval to the preceding or following maximum (M _1; M ₂ , ...) exceeds a predefined time threshold and / or - its respiratory movement measurement value (M _B ) exceeds one by the respective preceding or following maximum (M _1; M ₂ , ...) exceeds the specified threshold.

7. The method according to any one of the preceding claims, characterized in that a second period of low oxygen saturation (M ₀₂ ) is determined by determining maxima in the course of oxygen saturation (M ₀₂ ),

- after a time range of the drop in oxygen saturation (M ₀₂ ) is searched for the respective maximum,

- in which the duration of the drop in oxygen saturation (M ₀₂ ) is at least a predetermined time threshold, and

- At the end of the oxygen saturation (M ₀₂ ) by at least 3% compared to

The value of the maximum has decreased,

and such a time range is recognized as a second time period (T _Bi , T _B2 ,).

8. The method according to claim 7, characterized in that a maximum (M _1; M ₂ , ...) is only used as such if

- Its time interval to the previous or subsequent maximum (M _1; M ₂ , ...) exceeds a predetermined time threshold and / or

- Its saturation value exceeds a threshold value determined by the preceding or following maximum (M _1; M ₂ , ...).

9. A method for detecting a person's breathing interruptions,

- an oxygen saturation value (M ₀₂ ) indicating the relative oxygen saturation (M ₀₂ ) in the person's blood, and - In each case a feature vector (V _1; V ₂ , ...) from the measured values for a number of time intervals (l _1; l ₂ , ...), in particular of a duration between 0.5 and 5 seconds (M _b , M02) is determined,

- whereby from the respiratory movement measured values recorded in the relevant time interval (l _1; l ₂ , ...) a number of with regard to their position in the feature vector (V _1;

V ₂ , ...) predetermined feature vector entries is determined, and from the oxygen saturation values (M ₀₂ ) recorded in the relevant time window, a number of with regard to their position in the feature vector (V _1; V ₂ , ...) predefined feature vector entries are determined,

- A time window (F _1; F ₂ , ...) comprising a predetermined number of feature

Vectors (V _1; V ₂ , ...) or time intervals (l _1; l ₂ , ...) are specified,

the duration of the time window (F _1; F ₂ , ...) is selected such that it is preferably between 10 and 120 seconds,

- The feature vectors V ₁ assigned to the time window (F _1; F ₂ , ...) _; V ₂ , ... are transferred to a previously trained neural network (N) which contains in particular at least one recurrent layer, preferably a recurrent LSTM layer, and

- whereby the temporal sequence of individual probability values (P) is obtained by means of the neural network (N), which shows the probability of a misfire during the relevant time window (F _1; F ₂ , ...) or in relation to the relevant time interval (h , l ₂ , ...) for which a feature vector (V _1; V ₂ , ...) is valid.

10. The method according to claim 9, characterized in that the time windows (F _1; F ₂ , ...) overlap each other at least partially, in particular by half, so that for each by a feature vector (V _1; V ₂ , ...) characterized or affected by these affected two probability values (P _1; P ₂ ; P ₂ , P ₃ ) are determined, which indicate the probability (P) of a misfire at the relevant time, and that the two so for one Feature vector (V _1; V ₂ , ...) or a point in time obtained probability values, in particular weighted, are averaged.

1 1. The method according to claim 10, characterized in that the weighting of the separately determined and weighted added probability values (P _1; P ₂ ; P ₂ , P ₃ ) is predetermined such that values that are in the middle of the time window (F _1; F ₂ , ...) are weighted more than values that are at the beginning of the time window (F _1; F ₂ , ...).

12. The method according to any one of claims 9 to 1 1, characterized in that time ranges (B) in which the determined probability value (P) for the presence of misfires a predetermined, in particular by a moving average of the surrounding probability values, possibly multiplied by one predetermined factor, fixed, comparison threshold (P _th ) exceeds, as time ranges (B) are detected by breath failures.

13. The method according to any one of claims 9 to 12, characterized in that the neural network (N) for determining the probability value (P) has one or more further layers (N ₂ ) which are upstream or downstream of the LSTM layer and

- That the neural network (N) ends in a single neuron (N _out ), at the output of which the probability value (P) is obtained, or

- That the neural network ends in as many output neurons (N _out ) as there are time intervals in a time window, and at the individual outputs of the

Output neurons, the probability values assigned to the time intervals in the time window are obtained.

14. Data carrier on which a computer program for carrying out a method according to one of the preceding claims is stored.