WO2025042374A1 - Device for the diagnosis of obstructive sleep apnea - Google Patents

Abstract

The invention relates to a device for the diagnosis of obstructive sleep apnea.

Description

DESCRIPTION

DEVICE FOR THE DIAGNOSIS OF OBSTRUCTIVE SLEEP APNEA

Technical Field

The invention relates to a device for the diagnosis of obstructive sleep apnea.

Prior Art

The invention relates to a device for the diagnosis of Obstructive Sleep Apnoea (OSA). Sleep is an element that directly affects quality of life and OSA is a common disease characterised by respiratory arrest during sleep. The current diagnostic method, Polysomnography (PSG), is expensive, time consuming and uncomfortable for the patient. Furthermore, the diagnostic process can take a long time as there are not enough sleep laboratories and technicians. The present invention aims to develop a more practical, minimal in terms of sensors and apparatus, home-use device to overcome the difficulties of current methods used in the diagnosis of OSA. The device aims to provide early diagnosis of OSA by performing sleep staging and respiratory scoring accurately and effectively with signal processing and machine learning techniques [1].

Object and Brief Description of the Invention

The invention aims to identify sleep stages and respiration labels in the most efficient way in software. In this context, an artificial intelligence-based diagnostic algorithm has been developed using the Photoplethysmography (PPG) signal and the Heart Rate Variable (HRV) derived from this signal. In the invention, sleep staging and respiratory scoring processes were performed using the PPG signal and HRV features derived from 10 individuals. In statistical analyses, 75 features for sleep staging and 58 features for respiratory scoring were found significant. Classification was successfully performed with 84.93% sensitivity and 91.09% accuracy. The results show that the PPG signal and HRV can be used for sleep staging and respiratory scoring, so that these processes can be performed with a single signal. By processing the signals, the device calculates the Apnoea-Hypopnoea Index (AHI) and diagnoses OSA when AHI>5.

Figures

Figure 1. General flow diagram of the invention

Figure 2. Device wristwatch design

Figure 3. Portable handheld device design

Figure 4. Derivation of HRV parameter from ECG and PPG signals

Figure 5. Examples of filtering the photoplethysmography signal Figure 6. Moving average filter application

Figure 7. Derivation of HRV parameters from PPG signal Figure 8. Characteristic features of the PPG signal

Figure 9. Flow diagram for feature extraction from PPG signal

Figure 10. Detection of local minimum and maximum points of PPG signal and single period PPG signal Figure 11. General evaluation graph for classification results

1. Processor Device la - Electronic board section of the device. lb - Button for switching the device on and off.

1c - LCD screen.

Id - Input site of the pulse oximetry sensor. le - Speaker.

2. Sensor Cable

3. Pulse Oximeter

Detailed Description of the Invention

The operation of the invention is summarised in Figure 1. The device is designed in two different ways. The first one (Figure 2) is in wristwatch style and the second one (Figure 3) is portable. Both designs basically consist of three parts.

1. Processor Device

2. Sensor Cable

3. Photoplethysmography Signal Meter (Pulse Oximeter)

The functions of the parts for the wristwatch type (Figure 2.) are as follows. The part number 3 is the Pulse Oximeter. With this part, an electrical signal is measured according to the blood circulation in the finger. Many different values can be read at the same time. There are leds of different wavelengths at the sensor end. By means of these LEDs, information is obtained from the blood circulation at different wavelengths. In our study, operations can be performed with the information received at each wavelength. Part number 2 carries the information received from the oximeter sensor to the electronic device where the number 1 operations are performed. Part number 2 is the conductive cable. Part 1 is the electronic part for processing the information received from the sensor and making decisions. In a preferred embodiment, the device has a touch screen. By processing the incoming signal, it makes a decision about the disease by sleep staging and respiratory scoring.

The portable portable device (Figure 3.) is carried externally. In this design, the parts 2 and 3 in the system and their functions are the same.

The portable device shown in Figure 3 consists of 5 different elements, la is the electronic card part of the device and controls and operations are carried out through this card, lb is the button used to switch the device on and off. 1c is the LCD screen and the necessary information is displayed on this screen. The screen is touch screen. Id is the input section for transferring the information received by the pulse oximetry sensor to the device, le is used to inform the user through this speaker when the device gives a warning.

The records taken with the Pulse Oximeter Sensor are recorded by the device (1) during the night. These recordings vary according to the duration of sleep. The minimum measurement can be made over 30 seconds of recording. Two different operations are performed on the received signals. The first is sleep staging and the second is respiratory scoring. For sleep staging, the entire signal is divided into 30-second segments. Each part is labelled according to the machine learning methods designed in this invention. The labels are Sleep and Wakefulness. Together with the total sleep labelled segments, the time spent asleep is calculated (Equation 1). After this step is performed, the respiration scoring step is started.

In the respiratory scoring process, only 30-second segments labelled as sleep are checked. Whether there is a respiratory arrest in each segment is determined according to the machine learning methods realised in the present invention. The total number of respiratory arrests is determined. After that, the diagnosis of OSA is started. In this step, the total number of respiratory arrests is proportioned to the time spent asleep (Equation 2). This ratio is called the Apnea-Hypopnea Index (AHI).

OSA is diagnosed according to the AHI value obtained. The threshold value of AHI for the diagnosis of the disease is 5. If AHI<5, the individual is normal. If 5<AHI<15 is mild, 15<AHI<30 is moderate and 30<AHI is severe OSA.

Investigation of Photoplethysmography Signal and Heart Rate Variability

The PPG signal is a noninvasive, electrooptical method that provides information about parameters such as pulse rate, blood oxygen saturation and blood pressure depending on the volumetric change of blood flowing in a region of the body close to the skin [7], HRV is the analysis of time intervals of heart beats. Over a 24-hour period, the heart rate is continuously adjusted by the autonomic nervous system in response to internal and external triggers such as stress, rest, relaxation and sleep. With this change of HRV, it is widely used in the diagnosis of conditions such as sleep quality, OSA-related hypertension, sleep / OSA- related cardiac arrhythmias [3],

Generation and Measurement of Photoplethysmography Signal

PPG is a signal that can be measured from any part of the body. However, finger-based measurement devices are commonly used. The PPG signal is the basis of "pulse oximetry" technology [8], The PPG measurement setup basically consists of two parts. It consists of an LED, which is used to emit light of known wavelength (Z). and a phototransistor, which is designed to collect the remaining light after the tissues absorb the light. The position of the LED and phototransistor determines the name of the measurement mode. If the two parts are on the same side, this setup is called reflection measurement mode, if they are on different sides, it is called conduction measurement mode. Blood oxygen saturation measurement from PPG signal can be done by replacing the LED in Figure 4 with red and infrared LEDs [9].

Heart Rate Variable (HRV)

HRV is a parameter obtained by analysing the time between heart beats. This parameter is continuously adjusted by the autonomic nervous system depending on internal and external triggers such as stress, rest, relaxation and sleep.

Although HRV can often be derived from ECG signal, it has been proven that it can also be derived from PPG signal [10], Figure 4 shows the peaks of the ECG and PPG signals. HRV is the time elapsed between these peaks. The parameter calculated between each peak forms a component of the HRV sequence. The unit of HRV is second.

Digital Filtering

In the invention, MATLAB "fdatool" is used for digital filtering operations. A digital filter was designed and implemented to remove artefacts and noise on the PPG signal. Chebyshev Type II bandpass filter between 0.1 - 20 Hz and then moving average filter was applied to the PPG signal. Before and after filtering is shown in Figure 5 in 3 stages. The graph shows the unfiltered signal in the first stage. In the 2nd stage, only the Chebyshev Type II bandpass filter between 0. 1 - 20 Hz was applied to the signal. In the 3rd stage, the Chebyshev Type II bandpass filter between 0. 1 - 20 Hz and moving average filter are applied.

The reason for starting the filtering process at 0.1 Hz is to eliminate the DC components in the signal. In this way, DC components caused by absorption due to tissue absorption, absorption due to capillaries and absorption in the artery without heartbeats will be eliminated.

With MATLAB fdatool, 7 IR (infinite impulse response) and 11 FIR (finite impulse response) filters can be designed. Chebyshev Type II was selected among these filtering methods. Chebyshev filters can respond quickly to frequency transitions thanks to their steep descent in the extinction band. With the flat transition band, it can suppress unwanted frequencies better. Compared to other filters, the initial descent in the extinction band is sharper. Due to these advantages, this filter design has been used [11, 12],

Moving average filter is a simple filtering method for smoothing and removing small amplitudes in the signal. Similar to PPG and PPG, small amplitude signals in signals where peaks should be found cause false peaks to be found. In the invention, moving average filter is applied to the PPG signal to solve this problem. Figure 6 shows the moving average filter application. When R peaks were tried to be found without the moving average filter applied to the ECG signal, it was found that incorrect peaks were found. However, these errors were eliminated with the moving average filter.

Derivation of Heart Rate Variable

In Figure 7, graph 1 shows the 5-second PPG signal and graph 2 shows the HRV parameters derived from this signal. The sampling frequency of the PPG signal is fs=128 Hz and the total number of samples of the signal is 128x5=640. In order to derive the HRV, the local maximum points of the PPG signal were detected. The 4 local maximum points Api, AP2, AP3 and AP4 are marked with "*" in Figure 7. AHI, AH2 and AH3 values, which are the elements of the HRV sequence calculated using these points, are also marked with "*" in the graph number 2 in Figure 7. These marked points are the sample number corresponding to the point on the x-axis. For example, while the coordinates of the point where Api is located are [105 79.8726], API=105. While the coordinates of the point where AP2 is located is [246 80.5023], AP2=246.

Where N is the number of local maxima of the PPG signal and i=l,2,3,...,(N-l), the AH coefficients of the HRV are calculated as in Equation 3. The coefficients of the resulting HRV array are AH series. The number of elements of AH is N-l. An example calculation for AHI is shown in Equation 3.

Methods for Analysing the Photoplethysmography Signal and Heart Rate Variable

As with any biological signal, the PPG signal can be analysed according to its characteristic and statistical properties. By using these features, certain conditions of the body can be detected. For example, abnormal events occurring in respiration can be detected by looking at the amplitudes of the PPG signal [13], There are more than 20 characteristic features of the PPG signal in the literature [8], Figure 8 shows the most commonly used characteristics of the PPG signal in the literature. In addition to these characteristics, some statistical properties of the signal are used during the analysis.

The numbered features in Figure 8 can be explained as follows. Feature number 1 is the Systolic peak value, which is one of the characteristic features of the PPG signal. Dicrotic notch is another characteristic feature of the PPG signal. However, it may not be present in every signal. Codes that work to find this feature in real-time systems may cause errors. For this reason, it was not included in the study. Feature 2 represents the value of the bandwidth in seconds when the systolic peak amplitude is halved. Feature 3 is the value of the time in seconds from the Systolic peak amplitude to the point where the Systolic peak amplitude is halved. Feature 4 represents the time in seconds from the onset of a PPG signal to the Systolic peak amplitude . Feature 5 is the ratio of the A 1 and A2 areas to each other. This value is calculated as PA=A2/A 1. Feature 6 represents the time between two Systolic peaks. Feature 7 represents the time between the start and end of a PPG signal. All times are calculated in seconds. In addition to the characteristics of the PPG signal in Figure 8, the statistical properties of each period can be analysed. The mean, standard deviation, variance, minimum and maximum values of the signal can be considered as statistical features. It is possible to increase the number of these features.

After HRV parameters are extracted, the examination methods are similar to biological signals. As with any signal, feature extraction is performed. However, since there are not many characteristic features of HRV, statistical features are preferred.

The Data Set Used

In this section, the data set used within the scope of the invention study, the demographic information of the patients, the data sets allocated for sleep staging and respiratory scoring will be described.

Data Collection

The database was created from 33 channels of data recorded by the SOMNOscreen Plus brand PSG device in 10 individuals all night long in Sakarya Hendek State Hospital Chest Diseases Sleep Laboratory, with ethics committee approval and data usage permission. However, the study was performed only on the PPG signals received by the transducer placed on the abdomen. The sampling frequency of the data was 128 Hz and the patients slept for 7-8 hours. The extraction of PPG signal and HRV features and the feature selection algorithm will be detailed below.

Sleep staging data set

According to the sleep staging rules described, the available records were labelled by specialist Cahit BiLGIN. The sleep labels used were W, Nl, N2, N3 and REM. Since it is sufficient to determine the sleep and wake states for the respiratory scoring process, Nl, N2, N3 and REM were labelled with a single label as sleep (Sleep - S). In this way, two sleep labels, W and S, were used in total. The number of W and S stages obtained from each individual are shown. Each phase represents 30-second epochs. A total of 1482 awake epochs and 6953 sleep epochs were obtained from all individuals.

There are many ways to classify such unbalanced data. One of them is to select a sample from the excess data and balance it with the low number of data [14], In the results with unbalanced data, the accuracy rate of the results shifts towards the group with a high number and more errors are detected in the group with a low number [15,16], In order to avoid this problem that may occur in the invention, samples were selected from the large data set according to the systematic sampling theorem and the number of data was balanced. The data were balanced as 1482 (1481) and 6953 (1481).

Respiratory scoring data set Respiratory labels used in respiratory scoring are obstructive apnea, hypopnea, mixed apnea and central apnea. AHI calculated for the diagnosis of OSA is obtained by dividing the total number of apnea types by the sleep duration. According to this formula, it can be said that apnea types have no effect on the calculation of AHI index. In addition, in a study comparing obstructive apnea and hypopnea recordings, it was found that there was no statistically significant difference between apnea types [13], Therefore, in the present invention, all respiratory labels are combined under the general name of respiratory arrest. Thus, the respiration label to be given to the analysed signal is updated as "respiratory arrest" or "no respiratory arrest" .

For a recording to be labelled as "respiratory arrest", the recording must be at least 10 seconds long. The duration of the recordings in the study was at least 10 seconds.

Respiratory scoring data, male (1202) and female (1156) groups are close to each other in terms of numerical distributions. In this way, it can be easily seen that the groups are evenly distributed. In addition, apnea (1125) and control (1233) group data were also evenly distributed.

In order to carry out the study in a healthy way, a control group was formed against the respiratory labels. The respiratory label of the control group is "no respiratory arrest". The recordings taken for the control group were at least 10 seconds long as in respiratory arrests.

Feature Extraction

The feature extraction procedures to be described are common to both sleep staging and respiratory scoring. The only difference between the two groups is the length of the feature extracted signal. In sleep staging, recordings are divided into 30-second segments as standard. In respiratory scoring, the minimum recording time is 10 seconds. This can be up to 120 seconds. This difference in duration does not make a difference in the feature extraction process

Feature Extraction

A total of 46 features, 36 in time domain and 10 in frequency domain, were extracted from the PPG signal and 40 features, 30 in time domain and 10 in frequency domain, were extracted from the HRV. Since some of the features extracted from the PPG signal and HRV are common, a common expression was made in some places.

How the feature extraction process is performed from the PPG signal is shown in detail in the flow diagram in Figure 9. According to the flow diagram, the local minimum and local maximum points of the PPG signal are first detected. The signal is divided into periods according to the detected local minimum and local maximum points. The local minimum points of the signal are considered as the start and end points of the periods. The number of periods of the signal can be calculated as T=LOCMIN-1, where T is the period number of the signal and LOCMIN is the local minimum number. Figure 10 shows the detection of local minima and maxima of a 30-second PPG signal. Again, the first period of the signal, which is divided into periods according to the local minimum points detected in the same way, is shown. The number of periods of the 30-second PPG signal in Figure 10 can be found as T=LOCMIN-1=28-1=27 according to the local minimum number of the signal. The 30-second PPG signal in Figure 10 is divided into periods. The first feature to be extracted from this signal is obtained separately from the period according to the flow diagram in Figure 9 and averaged and saved as a feature of the 30- second PPG signal. This process is repeated for each feature extraction. For example, one of the features extracted from the PPG signal is the standard deviation of the signal. When calculating the standard deviation value, separate standard deviation values are calculated from the obtained period and the standard deviation value is averaged to obtain a single standard deviation value for the epoch. These steps are repeated for each feature extracted from the PPG signal. This process ensures that the features are extracted with minimum error rate.

Characteristic features of the photoplethvsmographv signal

In the literature, many different features have been extracted from the PPG signal [2, 4, 5, 6, 8, 13, 17-19], Some of these features are calculated depending on the shape of the signal [13], These are called characteristic features. In the invention, a total of 46 features were extracted from the PPG signal. The first 7 features extracted are characteristic features of the signal and are shown in Figure 10.

Statistical features of the photoplethysmography signal and heart rate variable

Statistical features are general parameter values obtained using descriptive measures of a signal of a given length. For example, the average of a 2-second ECG signal is a statistical feature of this signal. In the invention, some features of the PPG signal and HRV are extracted in common. Since the extracted features are the same, the expression will be common.

29 statistical features were extracted from PPG signal in time domain and 30 statistical features were extracted from HRV. Features 8-36 extracted from PPG signal in time domain and features 1-30 extracted from HRV in time domain are given in Table 1 with calculation formulae. Features marked with "*" were calculated with MATLAB [20], The x in the formulae represents the signal. Some features are shown as "- " in the "PPG/HRV Feature Number" column. For example, the "Normality Test p" feature is shown with in the "PPG Feature Number" column and with "27" in the "HRV Feature Number" column. This notation means that this feature is calculated for HRV, but not for PPG Signal.

HRV 26-27 are the features obtained as a result of Kolmogorov- Smirnov Normality Test. Kolmogorov- Smirnov Normality Test is one of the most common test methods used to test whether distributions are normally distributed [21], The p value obtained as a result of the test is the statistical probability value. If h=0, it represents the hypothesis Ho, and if h=l, it represents the hypothesis Hi. If p<0.05, h=l, and if >0.05, h=0. The features numbered 32-33 for PPG and 28-29 for HRV are the features obtained as a result of the sign test. Sign test is the nonparametric equivalent of the population mean significance test [21], It is a test that can be used when the population from which the sample is drawn is not normally distributed [21], This test was used because the single-period PPG signal did not show a normal distribution. Sign test tests whether the population median is equal to a certain value. The p value obtained is the statistical probability value.

If h=0, it represents the hypothesis Ho, and if h=l, it represents the hypothesis Hi. If p<0.05, h=l is determined as h=l, and if p>0.05, h=0.

In Table 1, the features 13-15 of PPG and 6-8 of HRV are extracted by the Hjorth method. Hjorth parameters are three different parameters derived to represent an x-signal in the time domain. These parameters are Activity, Mobility and Complexity parameters. In the formulae

, represents the variance of the x signal, , represents the variance of the 1st derivative of the x signal,

, represents the variance of the 2nd derivative of the x signal.

Energy level properties of the photoplethysmography signal and heart rate variable

In addition to statistical features, the energy amounts in the sub-frequency bands of PPG signal and HRV were determined and used as frequency domain features.

Table 1. Statistical time domain feature for PPG signal and HRV

While extracting the frequency domain features, the sub-frequency bands of the signals were first extracted. The sub-frequency bands of PPG and HRV signals are as shown in Table 7 [4, 18] .

In this study, PPG signal is divided into 3 different sub-frequency bands as low frequency band (LF), medium frequency band (MF) and high frequency band (HF). PPG signal is represented by PPG, LF band by PPGLF, MF band by PPGMF and HF band by PPGHF. HRV is divided into 3 different sub-frequency bands: very low frequency band (VLF), low frequency band (LF) and high frequency band (HF). HRV is represented by HRV, VLF band by HRVVLF, LF band by HRVLF and HF band by HRVHF-

Table 2. Ranges of sub-frequency bands of PPG and HRV signals

In order to obtain PPG and HRV sub-frequency bands, an IIR-Chebyshev Type II bandpass filter with the relevant frequency band ranges was designed and applied to the signals. After the application, 3 sub frequency bands were obtained from PPG and HRV. A total of 8 vectors were obtained with 6 subfrequency bands of PPG and HRV. The energies of these signals were calculated to calculate the frequency characteristics. Energy calculation was made according to Equation 4. Here x represents the signal whose energy is calculated.

The calculated energies are shown with the following symbols. PPG signal energy, PPG LF sub

frequency band energy, PPG MF sub frequency band energy, PPG HF sub frequency band

energy, E_HRV HRV energy, HRV VLF sub frequency band energy, HRV LF sub frequency

band energy, E_HRF HRV HF sub frequency band energy. The calculated features are numbered and shown in Table 3.

Table 3. Frequency domain features for PPG signal and HRV

Statistical Signal Processing. Feature Selection and Classification Algorithms Mann-Whitney U Test and Eta Correlation Coefficient were calculated for statistical analysis of the data. F-score feature selection algorithm was used as feature selection algorithm. For classification, kNN, MLFFNN, PNN, SVMs and ensemble classifier were used.

Performance Evaluation Criteria

Different performance evaluation criteria were used to test the accuracy rates of the proposed systems. These are accuracy rates, sensitivity, specificity, kappa coefficient (kappa value), receiver operating characteristic (ROC), area under an ROC curve (Area Under an ROC - AUC) and k-fold cross validation accuracy rate.

Sleep Staging

After statistical analyses of the features that can be used in sleep staging, they were classified with the help of classification algorithms. In addition to this process, 86 features extracted from PPG and HRV were selected and reclassified to improve the performance of the classifiers. F-score method was used for feature selection. F-score was applied to the features 2 times and the effect of the F-score method applied at different levels was analysed by classifying at each step. Table 4 shows the PPG and HRV features selected after the F-score method for sleep staging. In Table 4, the total number of features extracted from PPG and HRV is given in the number of features column. The features extracted in the column labelled "PPG HRV" were combined and used. The 46 features extracted from PPG were reduced to 21 features when the first F-score was applied. The feature numbers of these 21 features are shown in the column labelled "selected feature numbers". When F-score was applied to PPG for the second time, 21 features were reduced to 6. The same situation is also valid for HRV. A total of 40 extracted features were reduced to 14 when F-score was applied for the first time and to 6 when F-score was applied for the second time. When PPG and HRV were combined, a total of 86 features were reduced to 34 with the first F-score and 11 with the second one.

Table 4. Results of feature selection with F-score for sleep staging

32 35 37 38 25

After the feature selection process was completed, classification was performed. Classification processes were performed separately for PPG and HRV and then combined. In this way, all features were used. Firstly, 46 features of PPG were classified without any feature selection algorithm and performance parameters were calculated and written in the relevant column to measure the performance of the classifier. Then, 46 features were reduced to 21 with the first feature selection algorithm and the same process was repeated. Then, with the feature selection method for the second time, 21 features were reduced to 6 and classified. Then the performance criteria of the classifiers were calculated. The number of features obtained when each F-score is applied is also written. In the table, in the row named "Network Parameters", the network parameter information used for the relevant classifier is given.

The results obtained for the sleep staging process have been described so far. After this stage, the results of the respiratory scoring process will be presented.

The procedures for respiratory scoring were performed according to the flow diagram in Figure 1 and the results were obtained. The features extracted from PPG and HRV were analysed by statistical methods and it was investigated whether there was a significant relationship between them and respiratory labels. 58 of the 86 features of PPG and HRV were found to be significant. Non-significant features are marked in bold. As in the sleep staging process, since the PPG signal did not show a normal distribution, some descriptive parameters were included in the statistical analysis results.

After statistical analyses of the features that can be used in the respiratory scoring process, classification processes were performed. In addition, in order to increase the performance of the classifiers, 86 features extracted from PPG and HRV were selected and used with the F-score method. F-score was applied twice to the features and the effect of the F-score method applied at different levels was analysed by classifying at each step. Table 5. shows the PPG and HRV features selected after the f-score method for respiratory scoring. In Table 5, the total number of features extracted from PPG and HRV is given in the ‘Number of Features’ column. The features extracted in the column labelled ‘PPG HRV’ were combined and used. The 46 features extracted from PPG were reduced to 16 features when the first F-score was applied. The feature numbers of these 16 features are shown in the column labelled ‘selected feature numbers’. When F-score was applied to PPG for the second time, 16 features were reduced to 4. The same situation is also valid for HRV. A total of 40 extracted features were reduced to 11 in the first F-score application and to 5 in the second application. When PPG and HRV were combined, a total of 86 features were reduced to 28 with the first F-score and to 11 when applied for the second time.

Table 5. Results of feature selection with F-score for respiratory scoring

After the feature selection process was completed, classification was performed. Classification processes were performed separately for PPG and HRV and then combined. In this way, all features were used. Classification procedures were performed sequentially and the order is as follows. Firstly, 46 features of PPG were classified without any feature selection algorithm and performance parameters were calculated and written in the relevant column to measure the performance of the classifier. Then, with the first feature selection algorithm, 46 features were reduced to 16 and the same process was repeated. With the 2nd time feature selection method, 21 features were reduced to 4 and the classification process was repeated. The performance criteria of the classifiers are calculated and shown in the table . The number of features obtained when each F-score is applied is also written. In the tab named “Network Parameters” for each classifier, the network parameter used for the classifier was calculated.

Table 7 General evaluation table for classification results

The overall evaluation table for the classification results is shown in Table 7 and the graph is shown in Figure 11. According to Table 7, the best performance for sleep staging is obtained with the ensemble classifier. For this process, only HRV can be used or PPG and HRV can be used together. The combination of PPG and HRV both improved the classification performance and reduced the number of features, which is an advantage. The best performance for respiratory scoring was achieved with the ensemble classifier using PPG signal features. This process can be performed with both 16 and 4 PPG features. When 4 features were used in the system, it reduced the workload, while when 16 features were used, performance improvement was realised despite the increase in workload. It is possible to develop a physical system according to the obtained sleep staging and respiratory scoring results.

The second major aim of the invention is to provide reliable and practical detection of abnormal respiratory events occurring in the patient during sleep. For this purpose, respiratory scoring, which is the second step of the diagnosis, has been tried to be determined by the most practical methods.

The best results obtained from this invention are summarised in Table 7. Using the ensemble classifier and 11 PPG HRV features for sleep staging, a practical diagnosis system can be realised with an accuracy of 91.09%. If only HRV features are used, the number of features to be extracted will increase. In this case, the system can be realised with an accuracy rate of 90.01%. By using only HRV, practical sleep staging can be performed over the records in holter devices. In this way, the process of adaptation to hardware can be improved with only software changes.

Respiratory scoring can be performed with 93-95% accuracy rate only with PPG recording. The system can be realised with 16 PPG features and 95% accuracy rate. If savings are desired in terms of code writing, the number of features can be reduced to 4. In this case, the success rate will be 93.81%. This success rate is quite sufficient for the system to work. Considering all these values, it is possible to realise a practical sleep staging and respiratory scoring system. Figure 11 shows a graphical summary of the numerical values summarised in Table 7.

The ensemble classifier provided a superior performance in classifying each group of data. The reason for this is that the other classifiers compensate for the mistakes made by one classifier while performing individual classification. In this way, the power of the system is increased by combining individual performances. In addition, the distributions of the data are designed in such a way that there is no difference between the groups. The fact that the distributions are normal and uniform has a positive effect for the system to work healthier [22] .

REFERENCES

[1] Berry RB, Budhiraja R, Gottlieb DJ, et al, Rules for scoring respiratory events in sleep: update of the 2007 AASM Manual for the Scoring of Sleep and Associated Events. Deliberations of the Sleep Apnea Definitions Task Force of the American Academy of Sleep Medicine. J Clin Sleep Med 8:597-619. doi: 10.5664/jcsm.2172, 2012. [2] Kavsaoglu AR, Polat K, Bozkurt MR, A novel feature ranking algorithm for biometric recognition with PPG signals. Comput Biol Med 49: 1-14. doi: 10.1016/j.compbiomed.2014.03.005, 2014.

[3] Kim MS, Cho YC, Seo S-T, et al, Comparison of heart rate variability (HRV) and nasal pressure in obstructive sleep apnea (OSA) patients during sleep apnea. Measurement 45:993-1000. doi: 10. 1016/j .measurement.2012.01.044, 2012.

[4] Dehkordi P, Garde A, Karlen W, et al, Sleep stage classification in children using photoplethysmogram pulse rate variability. In: Comput. Cardiol. Conf. IEEE, pp 297-300, 2014.

[5] Gaurav G, Mohanasankar S, Kumar VJ, Apnea sensing using photoplethysmography. In: 2013 Seventh Int. Conf. Sens. Technol. IEEE, pp 285-288, 2013.

[6] Karmakar C, Khandoker A, Penzel T, et al, Detection of respiratory arousals using photoplethysmography (PPG) signal in sleep apnea patients. IEEE J Biomed Heal Informatics 18: 1065-1073. doi: 10.1109/JBHI.2013.2282338, 2014.

[7] Bal U, Bal A, Temassiz Fotopletismografi ile Nabiz Kestiriminde Hemoglobin Seviyesinin Etkisi. DOKUZ EYLUL UNlVERSlTESl MUHENDiSLiK FAKULTESl FEN VE MUHENDlSLlK DERGISi 17:47-53, 2015.

[8] Alian AA, Shelley KH, Photoplethysmography. Best Pract Res Clin Anaesthesiol 28:395-406. doi: 10.1016/j.bpa.2014.08.006, 2014.

[9] Bailey J, Fecteau M, Pendleton NL, WIRELESS PULSE OXIMETER. WORCESTER POLYTECHNIC INSTITUTE, 2008.

[10] Jeyhani V, Mahdiani S, Peltokangas M, Vehkaoja A, Comparison of HRV parameters derived from photoplethysmography and electrocardiography signals. Conf Proc . Annu Int Conf IEEE Eng Med Biol Soc IEEE Eng Med Biol Soc Annu Conf 2015 : 5952-5. doi : 10.1109/EMBC.2015.7319747, 2015.

1 1 1 1 Ucar MK, Bozkurt MR, Polat K, Bilgin C EEG Sinyalleri Kullamlarak Uyku Evrelerinin Simflandinlmasinda Sayisal Filtrelemenin Etkisi. In: Eleco 2014 Elektr. - Elektron. - Bilgi. ve Biyomedikal Miihendisligi Sempozyumu. Bursa, pp 27-29, 2014.

[12] Yavuz O, Can Bayram MC, Yddinm T, Chebyshev Filtre Parametrelerinin Yapay Sinir Aglan Kullamlarak Hesaplanmasi, Elektrik-Elektronik-Bilgisayar Miihendisligi 12. Ulusal Kongresi, pp 22- 25, Eskischir. 2007.

[13] Ucar MK, Bozkurt MR, Polat K, Bilgin C, Investigation of effects of time domain features of the photoplethysmography (PPG) signal on sleep respiratory arrests. In: 2015 23nd Signal Process. Commun. Appl. Conf. IEEE, pp 124-127, 2015.

[14] Ucar MK, Bozkurt MR, Bilgin C, Polat K, Automatic sleep staging in obstructive sleep apnea patients using photoplethysmography, heart rate variability signal and machine learning techniques. Neural Comput Appl 1-16. doi: 10.1007/s00521-016-2365-x, 2016.

[15] Perez-Godoy MD, Rivera AJ, Carmona CJ, del Jesus MJ, Training algorithms for Radial Basis Function Networks to tackle learning processes with imbalanced data-sets. Appl Soft Comput 25:26- 39. doi: 10.1016/j.asoc.2014.09.011, 2014.

[16] Rasch D, Teuscher F, Guiard V, How robust are tests for two independent samples? J Stat Plan Inference 137:2706-2720. doi: 10.1016/j jspi.2006.04.011, 2007.

[17] Elgendi M, On the analysis of fingertip photoplethysmogram signals. Curr Cardiol Rev 8: 14-25. doi: 10.2174/157340312801215782, 2012.

[18] Shi P, Zhu Y, Allen J, Hu S, Analysis of pulse rate variability derived from photoplethysmography with the combination oflagged Poincare plots and spectral characteristics. Med Eng Phys 31:866-71. doi: 10.1016/j.medengphy .2009.05.001, 2009.

[19] Addison PS, Respiratory effort from the photoplethysmogram. Med Eng Phys 41:9-18. doi: 10. 1016/j.medengphy.2016.12.010, 2017.

[20] Pascal Wallisch, Michael E. Lusignan, Marc D. Benayoun, Tanya I. Baker ASD and NGH, MATLAB for Neuroscientists: An Introduction to Scientific Computing in MATLAB. 550, 2014.

[21] Ramachandran KM, Tsokos CP, Mathematical Statistics with Applications in R. Math Stat with Appl R. doi: 10. 1016/B978-0- 12-417113-8.00006-0, 2015.

[22] Duan L, Xie M, Bai T, Wang J, A new support vector data description method for machinery fault diagnosis with unbalanced datasets. Expert Syst Appl 64:239-246. doi: 10.1016/j.eswa.2016.07.039,

Claims

1. A device for the diagnosis of Obstructive Sleep Apnea (OSA) disease, characterised by comprising a Pulse Oximeter for measuring the photoplethysmography signal (PPG) and by being configured to derive the Heart Rate Variable (HRV) from the obtained Photoplethysmography signal.

2. A device according to claim 1, characterised by comprising a processor (1) and a sensor cable (2) providing data flow between the processor (1) and the pulse oximeter.

3. A device according to claim 2, characterised by the processor (1) being wearable on the wrist of the user.

4. The method of operation of the device according to claim 1, characterised by a. Measuring and recording the photoplethysmography (PPG) signal b. Splitting the PPG signal into 30-second segments c. Deriving the Heart Rate Variable (HRV) by detecting the local maxima of the 30-second PPG signal d. Obtaining Sub Frequency Bands of PPG and HRV signals e. Determining characteristic and statistical features related to sleep and wakefulness f. Identifying sleep and wakefulness according to the characteristics g. Determining the characteristic and statistical properties associated with respiration h. Identifying respiratory arrests according to the characteristics for the sleep stages

5. A method according to claim 4, characterised by obtaining sub-frequency bands by filtering the PPG and HRV signals for determining characteristic and statistical features.

6. A method according to claim 4, characterised by separating the PPG signals into a low frequency band of 0.04 - 0.15 Hz, a medium frequency band of 0.09 - 0. 15 Hz and a high frequency band of 0.15 - 0.6 Hz.

7. A method according to claim 4, characterised by separating the HRV signals into a very low frequency band of 0.0033 - 0.04 Hz, a low frequency band of 0.04 - 0. 15 Hz, a high frequency band of 0.15 - 0.4 Hz.

8. A method according to claim 4, characterised by selecting sleep and wakefulness-related features by applying F-score and then selecting them a second time by applying F-score a second time.