Method and device for detecting respiratory disorders

(42) Date of commencement of the patent: 15. 9.1998 (45) Date of issue: 25. 8.1999

<td>(56) Documents:</td><td>(73) Patent owner:</td>

<td>US 5060279A US 5092343A</td><td>PFÖTZNER HELMUT DR. A-5640 BAD GASTEIN. SALZBURG (AT). nrrcruTif ΐέΆβι πρ A-2344 MARIA-ENZERSDORF, LOWER AUSTRIA (AT).</td>

(54) METHOD AND APPARATUS FOR DETECTING ANTI-FERTILIZER (57) For the detection of respiratory diseases, in particular sleep apnea, a method using physiological sensors and an adaptive signal analyzer is given. It is characterized in that a single detector is mounted on the body of the patient, which includes a plurality of sensors which operate nonspecifically and provide a series of mixing signals. The full information content of the mixed signal is used by being separated by signal splitters and the signals thus obtained are attributed to a family of simple characteristic images which are applied to an adaptive trained diagnostician trained diagnostics output.

CQ

AT 405 482

DVR 0078018

AT 405 482 Β

The invention relates to a method and a device for the detection of respiratory diseases, in particular sleep apnea, using physiological sensors and an adaptive signal evaluator.

For the diagnosis of respiratory diseases, physiological parameters which are directly related to the function of the lungs are conventionally used, such as the ventilation measured by means of spirography or plethysmography or the lung sounds detected by means of a stethoscope or microphone. To narrow the diagnosis, however, it is advantageous to also detect other parameters, in particular those that characterize the cardiovascular system. This is especially true in the case of sleep apnea. This term is understood to mean respiratory arrest occurring during sleep, in which case, in the case of obstructive apnea, the respiratory movements are essentially preserved, while in the case of central apnea they are lost. High medical relevance arises in the case of apnea of ​​the adult, which occurs especially in snoring persons, as well as in the case of infant apnea associated with SIDS.

Concerning physiological utterances, there are very similar problems in the case of other respiratory diseases. Particularly with regard to asthma diseases, there are widespread efforts to establish automatic diagnostic aids, whereby it is also a matter of inferring from the totality of several parameters on the type or degree of the disease. Because of the analogy of the technical implementation, only the concrete example of sleep apnea will be discussed below.

Complete diagnoses of apneas are usually based on the application of so-called polysomnography. The patient is monitored in the sleep laboratory for a full night by recording a large number of physiological parameters. However, the availability of such facilities is very limited even in leading industrialized countries. As an alternative, portable monitoring devices that can be used on normal hospital beds. The number of detected parameters is usually limited to four to eight, which is classified despite limited diagnostic value as an acceptable compromise (see. ASDA Standards of Practice in Sleep 17: 372-377, 1994). This can be an effective relief of Polysomnographiebetten achieve.

In order to reduce the high costs of inpatient examination, portable monitoring devices are also used for home monitoring. However, the corresponding efficiency of existing devices is limited due to lack of reliability. Previous device designs that capture only one or two parameters show good practicality, but diagnostic unreliability because of their low supply of physiological information. Devices which record the four to eight parameters already mentioned, which are generally classified as sufficient, prove to be problematic in home use because of their difficult operability. For the specific detection of the individual sensor signals, a plurality of detectors which are to be placed distributed over the body, for example ECG electrodes and stretch belts in various thoracic and abdominal regions, flow sensors in the mouth / nose area, but also oxymeter on the extremities (cf. eg Chest 108, 388-393, 1995). Malfunctions can thus result from the untrained layman not placing the detectors correctly or from the fact that the placements change during the monitoring. Ultimately, significant impairments to sleep quality can also result.

The regional distribution of the various sensors results from the general claim to register the individual parameters in a physiologically defined manner, as determined by the polysomnography or general clinical practice. For a specific parameter, the sensors are mounted so that the physiological size is optimally detected in a specific way. The specific, optimal detection of a physiological size has the advantage that the full information content can be used. So can eg from the signal of the impedance cardiography in a quantitative manner on the blood distribution of the heart are closed (see. eg WG Kubicek et al. Biomed.Eng.9, 410-416, 1974), but the electrode system is to be arranged parallel to the body axis. Likewise, a quantitative calibration can be achieved with respect to the ventilation of the lungs, provided the electrode system is placed normal to the body axis (K.Futschik et al. Ber. 14.J.Th.Ost.Ges.Biomed.Techn., 210-213, 1989). As another example electrocardiography may be mentioned in which a full diagnostic use of the signal waveforms presupposes compliance with one of the numerous, specifically standardized electrode placements.

A consideration of these conditions ultimately leads to the already mentioned complexity of currently common monitoring devices. As an example, the device described in EP 0 504 945 A2 is discussed in more detail. The positioning of the transducers on the body of the patient provides the following: In the upper and lower edge area of ​​the upper body, three disposable electrodes for detecting the ECG,

AT 405 482 Β a position sensor in its center, a microphone on the larynx for the detection of snoring sounds, an oxymeter sensor on one finger. In many devices, a flow sensor is additionally provided in the mouth / nose area. Due to this complicated detection system, efficient applications can only be achieved if the patient is enrolled at the clinic, or already provided there with specifically positioned sensor elements. Even with this effort but is expected to significant loss of data, which is encountered in some devices by two-sided modem communication (see. eg DPWhite et al. Sleep 18, 115-126, 1995).

An essential feature of the specific detection of a physiological size is to clean up the signal of artifacts. Thus, for example, when detecting the sleep sounds usually a microphone is used, which is attached directly to the larynx and has a frequency response that is specifically adapted to the snoring, whereby the latter can be detected by means of a threshold detector. Similarly, a thorax extension belt is usually placed and the signal supplied by him prepared so that the respiratory movement can be displayed as possible artifact-free.

The object of the present invention is to provide a method and a device for a monitoring device that is characterized by the simplest handling by the layman and hardly affects the quality of sleep.

According to the invention, this goal is achieved with that only a single detector is attached to the body of the patient, containing several different types of sensors, which do not work specifically and provide a series of mixed signals, whose full information content is thereby used that the signals are first preprocessed by a signal conditioner, be digitized and recorded on a storage medium and then fed off-line signal separators and separated from these, and that then the signals thus obtained are a bevy of simple, be fed to these downstream characteristic formers, whose outputs are connected to an adaptive diagnostics which is trained to diagnose output.

The measurement of defined artifact-free physiological signals is therefore omitted here, rather signal artifacts are used as essential information sources. Preferably, a large number of signal characteristic formers detect the high level of hidden information that is also contained in conventionally measured signals, but is usually discarded as an artifact. The signal characteristics are fed to the learning diagnoser, which is adapted to the individual unit of patient and detector as part of a night training.

The task of the detector is to collect such a large amount of physiological information at a single point of the body that it is sufficient for clear diagnosis. In the specific case of apnea detection, a distinction should be made - as a minimal requirement - between normal breathing, obstructive apnea and central apnea. One-point detection means a priori that the parameters measured by the detector are not detected in a specific way. If one were to specifically adapt the measuring point to a specific parameter, then it would be expected that all other parameters would be detected nonspecifically. According to the invention, a compromise is provided, according to which the detection takes place where there is an optimum for the totality of all parameters of interest.

According to the above, an oximetry measurement is ruled out a priori, since no further respiration-relevant parameters can be detected at the extremities. Furthermore, a flow measurement is not possible because there are no other parameters in the mouth / nose area. These two fundamentally very important parameters are thus eliminated a priori for a one-point measurement. According to the invention, the detector is positioned in a body region in which both the breathing activity and the heart activity can be detected, albeit only in an unspecific manner, in the sense of mixed signals. Figure 1 and Figure 2 show two possible variants of placements in which mixed signals can be measured, which proved to be sufficient for unambiguous diagnoses. These are placements in the heart or in the neck area.

Figure 1 shows a detector (1) placed close to the heart. It consists of a carrier strip (2) made of elastic, soft padded material and fully covered by an easily cleanable plastic skin, which can be attached to the skin, for example with two plaster strips. The exact placement can be marked on the clinic, the attachment can be made immediately before sleep by the patient himself. In the sketched embodiment, the plastic skin is interrupted only at four points for contacting four raised, made of stainless steel electrode cylinders (3), wherein the contact with the body is improved by Eiektrodenpasta. In addition, the carrier tape includes a solder sensor (4) and a miniature microphone (5), both of which are covered by the plastic skin. The detector (1) is via a single cable (6), the example in sleep interruptions by a clutch (7) can be interrupted, connected to the recording device (8), which for Signalaiabspeicherung with a Memo3

AT 405 482 B

Card writer is provided.

The mixed signal supplied by the microphone (5) consists of three contributions, which contain very extensive information. A contribution results from the heart activity, whereby the heart noises beside the directly relevant for apneas heart rate f_H is known to characterize a large number of cardiac and vascular anomalies. A second contribution from the lung activity characterizes the respiratory rate f, which approaches zero in central apnea_L, The signal pattern shows a priori strong individual and diagnostically usable fluctuations. Above all, it changes significantly in the case of obstructive apnea, which is used in the present method as well as the signal pattern of the third contribution, the snoring sound.

Also, the electrode cylinder (3) are used multiple times according to the invention. According to the known basic principle of electrical plethysmography in the heart region by means of the outer electrodes a high-frequency field (eg with 100 kHz). By means of the internal electrodes, a mixed signal is registered, the envelope of which is composed of three contributions used: The very low-frequency component characterizes the respiratory activity and delivers the lung frequency f, which approaches zero in central apneas_Lwhile obstructive apneas are manifested by reduction in amplitude and concomitant increases in harmonic content. A higher frequency component provides the heart rate, as well as - with limitations - information about temporal changes in the blood distribution of the heart. Ultimately, two electrode cylinders are also used to derive a nonspecific ECG signal, which also delivers the heart rate and, with restrictions, could also be used to detect extrasystoles that are very relevant to sleep monitoring.

The Lotsensor (4) consists in the case outlined of a cylindrical tube, on the inner wall of a molded soft magnetic material ball rolls, which is exposed to the magnetic field of a tube attached to the permanent magnet and their location from the - registered by means of a second attached field sensor - corresponding field distortion is derived. The height of the sensor signal encodes the body position essential for the apnea syndrome, the time derivative characterizes body movements and unrest.

Figure 2 shows a neck detector (9). Also, it consists of an elastic, soft padded and fully covered by a plastic skin carrier tape (10), but here preferably closed to a ring (eg through a Velcro tape). Again, a Lotsensor (11) is provided which provides substantially the same information as stated above. A microphone (12) provides a composite signal, in which case the snoring sound is dominant, followed by lung noise and heart murmur. The method of electrical plethysmography proved to be of little efficiency in testing the neck detector. According to the invention, a plethysmography sensor (13) operating according to known basic principles of the expansion principle is provided instead, for example a strain gauge mounted on an elastic carrier tape. It is placed over the carotid artery, providing a mixed signal whose low-frequency component characterizes the respiratory activity and the higher-frequency component the cardiac activity.

As with known devices, the sensor signals are preprocessed in the portable, preferably battery-powered recording device and preferably recorded after digitization. At the clinic, the signals are read out by means of a data reader and processed by a computer. According to the invention, this is done in three steps, the separation of the mixed signals, the formation of signal characteristics and the evaluation by the learning diagnostic tool.

For the separation of the mixed signals known methods of filtering are used, preferably the adaptive method described in AT 401340B. The low-frequency lung signal s_L and the higher-frequency heart signal Sh Essentially obtained by two filters, wherein in the first case, the upper and in the second case, the lower limit frequency is slightly below the heart rate f_H is set. The latter can be obtained in a simple, known way from the ECG signal in the case of the near-to-heart detector. In the neck detector, this adaptive approach is not possible because of the unavailability of an ECG signal. Experience has shown that the heart signal is relatively strongly contained here in the mixed signal of the strain sensor, whereby the signal separation in comparison to the near-heart plethysmography fails uncritical. Experience has shown that there are good separations in adult patients with cut-off frequencies around 0.7 Hz, resp. correspondingly higher values ​​in infants.

The procedure described above can in principle also be used to separate the acoustic mixed signals supplied by the microphone, which proves to be unequally more difficult here because of overlapping spectral components. A better known method which can be used in the present case is the separation with correlators using higher-order statistics. However, instead of an exact separation, it is also expedient for classification purposes to achieve the mixed signal by a series of inexpensive filters of Gaussian Durchiaßcharaketeristik staggered center frequency spek4

AT 405 482 Β traluminous coarse - in an unspecified way, according to AT 401226B - separate. The experience is used, according to which obstructive apneas are characterized by an increased occupancy of high spectral components. They are also associated with changes in respiratory and cardiac sounds that, although unspecific, can also be used diagnostically by a trained evaluator.

With the exception of f_H - and with restrictions also from Sl and s_H - The signals obtained in this way no strictly definable physical or physiological importance, but according to AT 401226B for classifications made by means of adaptive diagnostic agents, in particular neural networks, no significant disadvantage. As indicated in AT 401226B, the further signal processing can be carried out in a not strictly defined manner by the signals are supplied to several characteristic formers, which are designed as a simply constructed filter or signal distortion and take over the task of each signal more - as different, and to provide a good characterizing image. It makes sense to make a temporal averaging over the latter in order, for example, to obtain parameters that stand for time windows of 3 s duration. The latter are finally connected to the input of the diagnostician. Since apnea events generally have at least a duration of approximately 10 s, it is advantageous to supply the diagnosis builder with a sequence of parameters as input information, that is to say, for example three flocks of three consecutive time windows, ie the information of a total of 9 s.

The Diagnosebildner is constructed according to the prior art, such as a three-layer, supervised trained Neural network. The output may consist of a single neuron whose value represents the apnea event type, but each value may also be assigned its own neuron. The procedure described results in high numbers of parameters and thus also of input neurons. This can result in very high synapse numbers. The resulting requirement of very extensive training material is hardly a disadvantage here, since the latter is readily available in the case of apnea events. In addition, however, the availability of increasingly faster computers can result in the disadvantage of large computing times. It can be counteracted by the fact that the network is initially pre-trained for all characteristics, with known methods of network reduction (eg Pruning) but all those parameters are eliminated, which contribute little to a successful classification. The performance of the network thus trained on the basis of signal patterns of several patients may prove to be inadequate in the actual practical application to new patients, since individual physiological fluctuations and deviations of the detector placement can lead to greatly different signal patterns. According to the invention, a remedy is to nachzutrainieren the network based on individual signal pattern by the network a larger number of time windows for which the presence of apnea events can be excluded - eg from the time range before falling asleep - be classified as normal classified. Night training, which adapts the network to the individual entity formed by the patient and detector, is initiated automatically. However, the user can be asked to specify the time of falling asleep.

Figure 3 shows a possible embodiment of the signal processing. In the portable recording device, the signals coming from the sensors are preprocessed and digitized by signal conditioners (14). Thereafter, the signals thus obtained - the plethysmography signal s_Pi, the microphone signal Sm / c and the pilot sensor signal s_Loi - recorded on a memorized memory card (15). To get an ECG signal s_ekc, the signal delivered by the two inner plethysmographic electrodes is supplied to a low-pass filter (16) while the high-frequency plethysmographic component is decoupled by a high-pass filter (17).

The evaluation system is implemented on a PC. The four signals read offline from the memocard (15) are separated by filters (18) - (24). From s_P| becomes through a low-pass filter (18) the respiratory activity characterizing lung signal s_L obtained, by a high passport (19) the cardiac signal characterizing the blood release of the heart_H- Both filters are adaptively designed by the cutoff frequencies of the heart rate f_H be guided. For the safe determination of f_H will s_ekg used by the period is determined by means of the frequency detector (25) from the ECG main jaw spacing. From s_Mj_C by three simple Gaussian filters (20) - (22) three different imaging signals s_M1, s_MSecond s_M3 won. From s_Lot becomes a signal value SL through a low-pass filter (23)_age won, which codes for the body position, by a high pass (24) a signal s_Mot indicating body movements. The thus obtained eight partial signals are each supplied to three characteristic formers (25). The latter, in a rough approximation, determine the peak values, the RMS values ​​and the averages for time windows of 3 s - they can be constructed simply, since there are no demands for exactness; to demand is only uniqueness of the picture. To compensate for individual signal differences, these values ​​derived for 3 s are based on their respective long-term - eg determined for one hour - mean value. In the case of f_H

AT 405 482 Β, the task of the characteristic generator only refers to the averaging and the reference. This results in a total of 22 parameters that are applied to the input of a neural network (26). In order to provide the network with information about three consecutive time windows, it is also supplied with the intermediate memory 27 corresponding to the preceding time window, and also with the intermediate memory 28 which delays by 6 s. Overall, this results in the high number of 66 input neurons, which can be reduced by pruning. The network output consists of three neurons according to the classifications normal breathing (NA), obstructive apnea (OA) "and central apnea (ZA).