WO2009079977A2

WO2009079977A2 - Method and device for a preliminary sleep medical diagnosis and therapy monitoring

Info

Publication number: WO2009079977A2
Application number: PCT/DE2008/002034
Authority: WO
Inventors: Guy Leonard Kouemou
Original assignee: Guy Leonard Kouemou
Priority date: 2007-12-21
Filing date: 2008-12-04
Publication date: 2009-07-02
Also published as: WO2009079977A3; DE102007063007A1; WO2009079977A4

Abstract

The invention relates to a method and device for a preliminary sleep medical diagnosis and therapy monitoring. According to the method and the device for sleep phase detection, the sleep phases are subdivided into three phases deep, shallow and not-shallow or into two sleep phases deep and not-deep following a respiratory sound analysis. The method of the invention is characterized by recording respiratory sounds by means of a microphone system and using the same for sleep phase detection according to a neuronal spectral mean and RMS method.

Description

Method and device for sleep medical preliminary diagnosis and

Therapy accompaniment

The invention relates to a method for the diagnosis of sleep disorders in the home user area, based on the analysis of sleep-related breathing noises of the patients.

State of the art

Sleep has always aroused the interest of man. Today, a person spends about one-third of his life in sleep. The remaining two-thirds are also strongly influenced by the quality of sleep.

[0003] Therefore, especially in general medicine and sleep medicine many diseases can not only be prematurely diagnosed by an everyday observation of the quality of sleep. tiger recognized or better combated, but also avoided.

One of the focal points for evaluating the quality of sleep is, in addition to the sleep-related history (sleep-related questionnaire), the detection and analysis of sleep stages. In many sleep disorders, the duration of sleep stages 3 and 4 (deep sleep) is reduced or the deep sleep phases are not present at all.

A complex diagnosis of sleep disorders is currently possible only by a so-called. Electrophysiological Schlafpolygrafie that can be performed only in sleep laboratories of sleep medicine centers. Here, the patient is connected to a variety of sensors and examined for any existing sleep disorders. For the determination of the sleep stages are mainly the waveforms of the electroencephalogram (EEG), the electromyogram (EMG) and the Elektrookulogramms

(EOG) and evaluated according to the rules of Righteousness and Baldness.

Since such a diagnosis is very complicated and costly in a sleep laboratory, mobile so-called screening devices are generally used for the pre-diagnosis at home. These generally work on the basis of fewer sensors. From DE4114357C2 a device for the automatic detection of apnea is known. In this device, however, several sensors, including breathing belts in the chest, must be connected to the patient. This is often perceived as disturbing, since the sensors have to be worn during sleep. In addition, this device needs a separate reading and writing device for recording the data, which must be evaluated by a doctor afterwards.

From DE69930720T2 another monitoring device for the detection of sleep apnea is known. This device is incorporated in a small plastic unit which is positioned under the nose of the patient. The device measures the air flows during inhalation and exhalation and records the respiration or respiratory failure during sleep. There is a much simplified form of diagnosis based on the recorded data. A doctor can not rely on the recorded data to make his own diagnosis, so the method is only suitable for patients with little chance of real pathology.

Already for such a pre-diagnosis at home a doctor's visit is necessary, unless the patient invests itself in the purchase of a corresponding device. However, it is already necessary for a doctor's visit for the patient to be able to recognition as an illness. This costs experience, some overcoming, especially in the case of insomnia, the patient can not observe his own disease itself, but is usually dependent on statements of the bed partner.

task

The basic idea of the developed in this work, new method for sleep phase detection is based on observations of the author, who were supported by doctors and staff of the research group:

If one observes a sleeping person, one can estimate on the basis of the respiratory sounds or their variability, whether the person sleeps quietly or restlessly. Assuming a more even breathing in deep sleep phases (sleep stages 3 and A), an assessment of whether someone is in such a sleep, based on a breath sound analysis seems possible.

Although this hypothesis is neither proved nor disproved after an intensive literature search, the author decided to investigate the connections within the framework of a research work on the basis of the motivating aspects mentioned above.

The technical objective and the object of the invention is therefore to provide a practical and easy to use by any patient To develop a method with the associated device system or corresponding software that analyzes and displays the quality of sleep and thus provides the patient with rapid information about possible sleep disturbances or lack of deep sleep phases. In particular, attention should be paid to simple operability and low senorik, preferably only in the form of a microphone.

This object is achieved according to the invention using the method characterized in the claims with associated software. The core of the invention consists in the fact that the evaluation of the quality of sleep by the software using artificial neural networks is performed and the evaluation is performed only on the basis of a breath sound recording.

The classification is based both on the evaluation of the root mean square squares over a certain signal duration, as well as on the evaluation of spectral components of the breath sound signal.

Both RMS values and spectral components are calculated and stored online during the recording by the software. The classification with the aid of a neural network then takes place, wherein the previously described stored values are loaded and processed as described in the patent claims under number 1. Solution proposal and embodiment

Method for recording measured values and extracting the characteristic features:

The breath sounds are recorded using an airborne microphone or a structure-borne sound microphone. The airborne microphone is attached to a headband and attached to the patient's head (see Figure 2). The Acoustic Microphone can also be mounted on a stand at a distance of 0-50cm from the patient's head (see Figure 3). For the use of the Köperschallmikrofons this is preferably attached to the upper lip (see Figure 4) or to the trachea. With a simple sound blaster, breath sounds with a sample rate of, for example, up to 11.025 kHz and with e.g. 16 bit resolution sampled. In a subsequent first processing step, the data is compressed using a bandpass filter to a frequency range between 0 and 5500 Hz, since no higher frequencies are contained in the breath sounds.

After this preprocessing, for example, a time-dependent and partial spectral analysis is performed. For each short-term spectrum, 256 samples were used by way of example (about 10 ms because of short-term stationarity of respiratory sounds). To accelerate the subsequent pattern recognition becomes a Data reduction by a factor of up to 16 using a filter bank approximation. The spectrum is divided into eg 16 areas. The mean values of the respective areas then form the filter bank coefficients, which are then combined to form a new feature vector. The filter bank analysis produces a compact approximation of the power density spectrum describing the signal as a sequence of feature vectors. Thus one obtains the spectral mean values of the breath sounds.

Transferred to this case of investigation, it is determined by means of the correlation coefficient whether there is a relationship between the feature matrix and the polysomnographic sleep stage vector.

The analysis of the frequency banks showed that the four lower ones correlated better. For this reason, preferably the 4 lower frequency banks are included for sleep phase detection.

Example filter banks: Bankl (0-124 Hz), Bank2 (124-264 Hz), Bank3 (264-421Hz), Bank4 (421- 597Hz).

In parallel with the determination of the spectral values of the respiratory sounds in the frequency domain, the "root mean square" of the signal is calculated in the time domain.The root mean square can be understood in German as the effective value of the signal.

The RMS of a signal x out of n sample points is calculated as follows:

[0023] In accordance with the present method, it has preferably been calculated from the patient breath data of an RMS vector, each with n = 30x11025 = 330750 points (N.B., with segment duration of, for example, 30 seconds, sampling frequency, for example, 11025 points / second, mono).

Combination of the spectral mean values with the RMS:

The feature extraction consists of 2 stages. In the first stage, the data is normalized.

In the second stage, a 5 × N feature matrix is preferably calculated from the individual respiratory sounds. N denotes the single vector from the segment duration a 30 seconds per vector (preferably). This corresponds, e.g. 2 points per minute, i. 120 points per hour, or 1200 points for a night recording of 10 hours.

With the 5 features per vector (every 30 seconds), four indicate the amplitude of the 4 lower frequency ranges (ie, for example, 0-124Hz, 124-264Hz, 264-421Hz, 421-597Hz) of the spectral averages, and one the RMS amplitudes.

Segmentation process: The studies carried out in the present work were carried out against polysomnogramin derived sleep stages.

Respiratory noises of several patients during an entire night in an accredited sleep laboratory according to the recording method described above by means of a microphone system and a PC are recorded.

The recordings are performed for example with a self-written program. Above all, the program is characterized by its high stability and robustness when handling large amounts of data (up to more than 650 MB over 10 hours recording here). In addition, the program has some well-used analysis and signal processing functions that are very helpful for rapid viewing and analysis of the data.

The polysomnogram recorded in the sleep laboratory is normally evaluated automatically by the system analysis program the next day. In the evaluation, sleep phases and apnea appearances are calculated as a function of time by the system.

The responsible doctor later revised the automatic evaluations. Among other things, it checks the automatic determination of sleep stages carried out by the analysis program, and corrects them if necessary. A revision by the doctor can usually take between 45 minutes to over 2 hours to last; to endure, to continue. Since many patients in the sleep laboratory do not always have deep sleep phases due to sleep disturbances, it is important when examining the sleep phases to include both ill and healthy patients.

In the developed new method, for example, noise signals were processed in blocks of 30 seconds, since the sleep stages calculated from the polysomnogram used in the test were always determined over a period of 30 seconds.

Design of the neural recognizer:

First, a multilayer preferably forward-directed neural network is designed.

Then a training procedure with a corresponding algorithm (here preferably backpropagation) is defined. When defining the training process, network initialization function, segmented test and validation patterns, number of test cycles, termination criteria, and training duration should be entered. Then the training of the neural network is started and finally a network is generated. The trained network is then connected to the system in a modular way.

Description of the system The system consists of a microphone system, a PC with AD-DA converter card (sound blaster card) and a monitoring tool.

Figure 5 shows an exemplary schematic representation of the processes in the software. First, the recorded breath sound data is digitized. Subsequently, the data is filtered and normalized in the signal preprocessing. Next, the feature extraction and handover of the data to the neural network or Hidden Markov Model Network occurs. In this network, the detection of the sleep phases takes place. The network is trained with a teaching database containing recorded data from clinical sleep laboratories. These data are passed to another neural network or hidden Markov model network, where the detection of sleep disorders and sleep phase patterns takes place. In the process, the network is trained with a further instruction database containing typical sleep disorder pictures from clinical sleep laboratories. If sleep disorders are detected, an advance diagnosis with therapeutic suggestions is made using a knowledge-based automaton.

Claims

Patent claims

1. A method for detecting and visualizing the sleep profile by means of the analysis of sleep-related breath sounds, characterized in that for the detection of sleep-related breath sounds only a nocturnal ^" Atemgeräusch- Aufzeichnung, preferably using a mounted on a headband above the nose airborne microphone, takes place that the the signals recorded by the microphone are preferably fed to a preamplifier, that the amplified signals, digitized in a sound card, are made available to the program.

2. The method according to claim 1, wherein the digitized signals are normalized and filtered.

3. The method of claim 1 and 2, wherein from the thus processed signals segmentally over a certain time range, preferably an epoch, the effective values (Root Mean Squares) are calculated and stored.

4. The method of claim 1-3, wherein spectral components are calculated from the filtered and normalized signals and over a certain period of time, also preferably an epoch, gerαittelt and stored.

5. The method of claim 1-4, wherein the stored RMS values and the stored spectral components of the lower frequency ranges of the entire recording are loaded in the program and assembled in a matrix.

6. The method of claim 1-5, wherein the combined values are time-smoothed.

7. Method according to claims 1-6, the values are standardized to a mean of zero and a standard deviation of one.

8. The method of claim 1-7, wherein the values for the classification of sleep stages are fed to a previously trained, artificial neural network and evaluated by this.

9. Software program for carrying out the project according to claim 1-8.