Diagnosis of obstructive sleep apnea with respiratory function test

Abstract

The invention relates to a method regarding the diagnosis of OSA by means of extracting characteristic and statistical features from measurement curves relating to breath speed, intensity, and frequency during a respiratory function test (spirometry) performed on patients, and analyzing these features with artificial intelligence-based methods.

Description

[0001] DIAGNOSIS OF OBSTRUCTIVE SLEEP APNEA WITH RESPIRATORY FUNCTION TEST

[0002] Technical Field

[0003] The invention relates to a method regarding the diagnosis of obstructive sleep apnea (OSA) by means of extracting characteristic and statistical features from measurement curves relating to breath speed, intensity, and frequency during a respiratory function test (spirometry) performed on patients, and analyzing these features with artificial intelligence-based methods.

[0004] Prior Art

[0005] The respiratory function test (spirometry) is a physiological test administered for the purpose of evaluating the lung functions of individuals. In this test, the performance of the lungs is analyzed by measuring the flow and volume changes occurring during the person's breathing. The respiratory function test aims to measure how much air enters and exits the lungs, the speed of the air during respiration, and the effort capacity. The most widely used method among respiratory function tests is spirometry, and with this method, flow and volume values are recorded.

[0006] According to international recommendations, the diagnosis of OSA is made using scoring criteria after sleep examination with polysomnography (PSG). In diagnosis with PSG, the individual sleeps for a night in a sleep laboratory under expert control, connected to numerous electrodes on their body. The physiological information recorded together with video and audio recordings is subsequently examined by a specialist doctor. According to this; apnea is defined as a 90% reduction in airflow lasting at least 10 seconds, and Hypopnea is defined as at least a 50% reduction in flow for a period of at least 10 seconds and a 3% reduction in oxygen saturation. OSA severity is determined according to the apnea-hypopnea index (AHI) value, which is the number of apnea-hypopnea per hour of sleep. The AHI value is defined as; <5 no sleep apnea, 5-15 mild OSA, 15-30 moderate OSA, >30 severe OSA.

[0007] However, the PSG method carries significant disadvantages such as high cost, requirement for experts, being time-consuming, and reducing patient comfort. In the document Köktürk, Oğuz. "Obstrüktif uyku apne sendromu yardımcı tanı yöntemleri." Tüberküloz ve Toraks dergisi 48.1 (2000): 79-86., it is stated that respiratory function tests could provide supportive information in the diagnosis of OSA; however, it is emphasized that they are not direct diagnostic methods.

[0008] In the study performed by Seval Bulut Eriş and colleagues and published in the journal International Journal of Medical Engineering and Informatics with the title “A review of the relationship between flow-volume curve and obstructive sleep apnea” (DOI: 10.1504 / IJMEI.2023.10054791), the relationship between the flow-volume curve and obstructive sleep apnea was examined, and it was determined that there is a limited number of studies on this subject in the literature and that the studies conducted were mostly limited to the statistical analysis of ready spirometry parameters. Furthermore, it was emphasized that OSA diagnosis could be facilitated by developing machine learning-based systems with the use of new characteristic and statistical features that can be extracted from the flow-volume curve.

[0009] When the state of the art is examined, it is seen that no original study directed towards making an OSA diagnosis by extracting new characteristic features from flow-volume curve data obtained from the respiratory function test and analyzing these data with artificial intelligence methods has been encountered. It is understood that in existing studies, standard spirometry parameters are used directly, and raw flow-volume curve data are not processed with advanced analysis methods. In this context, the use of these features in artificial intelligence-based diagnostic systems by performing original feature extraction constitutes a significant novelty in the existing technical field.

[0010] Therefore, when the studies existing in the prior art are examined, a need has been felt for the development of obstructive sleep apnea diagnosis with respiratory function test.

[0011] Objectives and the Brief Description of the Invention

[0012] The object of this invention is to extract characteristic and statistical features of the values obtained from the changes in speed, intensity, and frequency during respiration as a result of the respiratory function test performed on many selected patients, and to develop an original method capable of being used in OSA diagnosis using these features.

[0013] Another object of this invention is to produce new attributes from test outputs instead of ready parameters obtained directly in traditional tests, and to enable these attributes to be analyzed with artificial intelligence algorithms. Another object of this invention is to perform OSA diagnosis with high reliability by means of the developed artificial intelligence-based diagnostic system, with performance values of 97.1% accuracy, 0.9744 sensitivity, 0.9672 specificity, 0.9744 precision, 0.9744 F-measure, 0.9694 kappa, and 0.9884 area under the receiver operating characteristics (ROC) curve (AUC).

[0014] Detailed Description of the Invention

[0015] The method for processing data directed towards obstructive sleep apnea diagnosis with respiratory function test, it comprises;

[0016] determining the start points of flow-volume curve maneuvers according to breathing in and out data,

[0017] obtaining separate flow-volume curves using the start points,

[0018] calculating the total lung capacity (TLC) value of each flow-volume curve by subtracting the start point volume value from the end point volume value, selecting the flow-volume curve having the highest TLC value for being processed in signal processing stages,

[0019] determining characteristic and statistical features from the selected curve, creating a database with the extracted features,

[0020] calculating the impact levels of the features in determining the class label using nonparametric tests,

[0021] creating different models using machine learning methods and feature selection algorithms,

[0022] evaluating the performance of the models with certain values,

[0023] determining whether the individual belongs to the obstructive sleep apnea class as a result of the evaluation.

[0024] The specified classification result regarding the presence of OSA assists the doctor in making a definitive diagnosis.

[0025] Among the records of those applying to the hospital, patients referred to PSG with an OSA clinical finding and / or complaint, and those coming for a check-up due to an OSA diagnosis, were examined retrospectively. In this context, 633 patients applied to the Department of Chest Diseases. As a prerequisite, the individuals to be included in the study as OSA patients are required to meet one of the following criteria; - being referred to PSG with a preliminary diagnosis of OSA and receiving an OSA diagnosis as a result of PSG,

[0026] CPAP / BPAP treatment being planned with an OSA diagnosis,

[0027] receiving CPAP / BPAP treatment with an OSA diagnosis.

[0028] 234 individuals meeting one of the criteria above and also having a PFT record were included in the study as OSA patients. Among the individuals having a PFT record, 232 individuals having no clinical finding and / or complaint regarding OSA, or who were referred to PSG with a preliminary diagnosis of OSA and whose PSG result turned out negative, were also selected as control data. The number of data was determined using the sampling theorem. The universe of the study is around 633 patients. Although OSA prevalence varies from society to society, studies conducted state that OSA prevalence in the general population is between 9% and 38%. When the sample size is calculated by taking the highest rate in the general population into account and taking the variance of the population as 38%, the confidence interval as 95%, and the margin of error as 5%, the number of patients was found to be 231. The amount of data used for the study to be performed is considered sufficient. Furthermore, when the study was repeated by increasing the number of data, no significant change was observed in the results.

[0029] In the application of the subject invention, the steps of:

[0030] determining the patients to participate in the diagnostic study,

[0031] obtaining flow-volume curves by performing respiratory function tests on the patients, extracting features from the obtained flow-volume curves,

[0032] calculating the correlation coefficients of the extracted features with the OSA class, ranking the features from high to low according to correlation coefficients and creating different data groups by taking this ranking into account,

[0033] creating classification models using 5 different machine learning methods for each data group,

[0034] performing performance evaluations of the models are included.

[0035] For the step of Obtaining Flow-Volume Curves; respiratory function test (PFT) records performed with computer-based microQuark spirometry in a sitting position were obtained using the Omnia version 1.6.10 software. Flow-volume curves were plotted in the Matlab environment from the obtained PFT records. In order to separate PFTs having more than one maneuver from each other, they were marked from the expiration end points. The total lung capacity (TLC) value of each flow-volume curve was calculated, and the flow-volume curve having the highest TLC value was selected. This selection method is different from the ATS / ERS 2019 Spirometry Standards international standards and is an original approach of the study.

[0036] In the feature extraction step; 94 characteristic features (Table 1) and 25 statistical features (Table 2) were extracted from the selected flow-volume curve, independently of the measurement results provided by the spirometry. MATLAB's ready-made codes were used for some of the statistical features. The other statistical features and all characteristic features were calculated by means of codes written in the MATLAB program.

[0037] Feature Feature Unit Description Feature Feature Unit Description Number Number

[0038] 1 TLC L Total Lung 48 PIF_FIF75time / Feature 36 / Feature Capacity. It is PEF_FEF75time 28

[0039] the air volume

[0040] in the lungs

[0041] with deep

[0042] inspiration.

[0043] 2 FVC L Forced Vital 49 TLC / FVC Feature 1 / Feature 2

[0044] Capacity. It is

[0045] the volume of

[0046] air exhaled with

[0047] rapid and

[0048] forceful

[0049] expiration

[0050] following deep

[0051] inspiration.

[0052] 3 FEV₁ L Forced 50 PIF / PEF Feature 14 / Feature Expiratory 5

[0053] Volume in the

[0054] first second

[0055] 4 FEV₁ / % Feature 3 / 51 FIF25 / FEF25 - Feature 15 / Feature FVC Feature 2 6 5 PEF L / sn Peak Expiratory 52 FIF50 / FEF50 - Feature 16 / Feature Flow 7

[0056] 6 FEF25 L / sn Forced 53 FIF75 / FEF75 - Feature 17 / Feature expiratory flow 8

[0057] at 25% of FVC

[0058] 7 FEF50 L / sn Forced 54 Area_54 L / sn Area between the expiratory flow start of expiration at 50% of FVC and PEF

[0059] 8 FEF75 L / sn Forced 55 Area _55 L / sn Area between PEF expiratory flow FEF25 at 75% of FVC

[0060] 9 PEFtime sn Time to PEF 56 Area _56 L / sn Area between FEF25 andFEF50 10 FEF25time sn Time to FEF25 57 Area _57 L / sn Area between FEF50 and FEF75 11 FEF50time sn Time to FEF50 58 Area _58 L / sn Area between FEF75 and the end of

[0061]

[0062] expiration FEF75time sn Time to FEF75 59 Area _59 L / sn Area between the start of inspiration and FIF25 FEF25 - L / sn Mean forced 60 Area _60 L / sn Area between FIF25 FEF75 expiratory flow and PIF between 25%

[0063] and 75% of

[0064] FVC PIF L / sn Peak 61 Area 61 L / sn Area between PIF Inspiratory and FIF50 Flow

[0065] FIF25 L / sn Forced 62 Area _62 L / sn Area between FIF50

[0066] Inspiratory and FIF75 Flow at 25% of

[0067] FVC FIF50 L / sn Forced 63 Area _63 L / sn Area between FIF75

[0068] Inspiratory and the end of Flow at 50% of inspiration FVC FIF75 L / sn Forced 64 Area 64 Ratio of the area Inspiratory between the start of Flow at 75% of expiration and PEF FVC to the area from PEF to the end of expiration PIFtime sn Time to PIF 65 Area _65 Ratio of the area between the start of inspiration and PIF to the area from PIF to the end of inspiration FIF25time sn FIF25 Time 66 Area 66 Ratio of the area between FEF25 and FEF75 to the area between FIF25 and FIF75 FIF50time sn Time to FIF50 67 Area _67 - Feature 55 / Feature 56 FIF75time sn FIF75 Time 68 Area _68 - Feature 56 / Feature 57

[0069] FIF25 - L / sn Mean forced 69 Area _69 Feature 60 / Feature FIF75 inspiratory flow 61

[0070] between 25%

[0071] and 75% of

[0072] FVC TLCtime sn TLC time 70 Area _70 - Feature 61 / Feature 62

[0073] FVCtime sn Time to FVC 71 TLC / PIF sn Feature 1 / Feature 14 FEF25_FE sn Time to 72 FVC / PEF sn Feature 2 / Feature 5 F75time FIF25_FIF75.

[0074] Time to 50% of

[0075] the FVC

[0076] maneuver.

[0077] PEF_FEF2 sn PEF_FEF25 73 TLC / sn Feature 1 / (Feature 5time (PEF + PIF) 5 + Feature 14) PEF_FEF5 sn Time to 74 FVC / sn Feature 2 / (Feature Otime PEF_FEF50 (PEF + PIF) 5 + Feature 14) PEF_FEF7 sn Time to 75 TLC / PEF sn Feature 1 / Feature 5 5time PEF FEF75

[0078] FEF25_FE sn Time to 76 FVC / PIF sn Feature 2 / Feature F50time FEF25 FEF50 14 PEF_FEF2 - Feature 26 / 77 PIF / FIF25 - Feature 14 / Feature

[0079]

[0080] 5time / Feature 27 15 PEF_FEF5

[0081] Otime

[0082] 31 PEF_FEF5 Feature 27 / 78 PIF / FIF50 Feature 14 / Feature Otime / Feature 28 16 PEF_FEF7

[0083] 5time

[0084] 32 FEF25_FE Feature 29 / 79 PIF / FIF75 Feature 14 / Feature F50time / Feature 25 17 FEF25_FE

[0085] F75time

[0086] 33 FIF25_FIF sn Time to 80 PEF / FIF25 Feature 5 / Feature 75time FIF25_FIF75. 15

[0087] Time to 50% of

[0088] the TLC

[0089] maneuver.

[0090] 34 PIF_FIF25t sn Time to 81 PEF / FIF50 - Feature 5 / Feature ime PIF FIF25 16 35 PIF_FIF50t sn Time to 82 PEF / FIF75 - Feature 5 / Feature ime PIF FIF50 17 36 PIF_FIF75t sn Time to 83 TLC / FIF25 sn Feature 1 / Feature ime PIF FIF75 15 37 FIF25_FIF sn Time to 84 TLC / FIF50 sn Feature 1 / Feature 5 Otime FIF25 FIF50 16 38 PIF_FIF25t Feature 34 / 85 TLC / FIF75 sn Feature 1 / Feature ime / Feature 35 17 PIF_FIF50t

[0091] ime

[0092] 39 PIF_FIF50t Feature 35 / 86 FVC / FEF25 sn Feature 2 / Feature 6 ime / Feature 36

[0093] PIF_FIF75t

[0094] ime

[0095] 40 FIF25_FIF Feature 37 / 87 FVC / FEF50 sn Feature 2 / Feature 7

[0096] 5 Otime / Feature 33

[0097] FIF25_FIF

[0098] 75time

[0099] 41 FIF25_FIF Feature 33 / 88 FVC / FEF75 sn Feature 2 / Feature 8

[0100] 75time / Feature 25

[0101] FEF25_FE

[0102] F75time

[0103] 42 PIF_FIF25t Feature 34 / 89 FIF25 / FIF50 Feature 15 / Feature ime / Feature 26 16 PEF_FEF2

[0104] 5time

[0105] 43 PIFtime / - Feature 18 / 90 FIF25 / FIF75 - Feature 15 / Feature PEFtime Feature 9 17 44 FIF25time - Feature 19 / 91 FIF50 / FIF75 - Feature 16 / Feature / Feature 10 17 FEF25time

[0106] 45 FIF5 Otime - Feature 20 / 92 FEF25 / FEF50 - Feature 6 / Feature 7

[0107] / Feature 11

[0108] FEF50time

[0109] 46 FIF75time - Feature 21 / 93 FEF25 / FEF75 - Feature 6 / Feature 8

[0110] / Feature 12

[0111] FEF75time

[0112] 47 PIF_FIF50t - Feature 35 / 94 FEF50 / FEF75 - Feature 7 / Feature 8 ime / Feature 27

[0113] PEF_FEF5

[0114]

[0115] Otime

[0116] Table 1: 94 characteristic features obtained from the flow -volume curve Featur Feature Equation Featur Feature Equation

[0117] e e

[0118] Numbe Numbe

[0119] r r

[0120] 95 Mean jA, 108 Harmonic Mean H=n / (1 / x_1 +...+1 / x_n)

[0121]

[0122] J=1

[0123] 96 Hjorth Parameters- Activity / Variance A=S² 109 Median

[0124] Hj orth Parameters-—Activity / Variance

[0125]

[0126]

[0127] 97 110 Minimum* xmin=min(xi) Standard Deviation

[0128]

[0129]

[0130] 98 Mean Curve Length 111 Maximum* Xmax=max(Xi)

[0131] 99 Mean Energy j " 112 Range R=xn-Xi

[0132] E = - ) A-‘

[0133] 100 Mean Teager Energy j _ 113 Singular Value

[0134] =Decomposition SVD=svd(x)

[0135]

[0136] 101- M=S12 / S2 11425% Trimmed T25=trimmean(x,25) Hjorth Parameters -1 7Mean*

[0137] Mobility

[0138] 102 115 50 / o Trimmed T50

[0139]

[0140] —trimmeanfx 50) Hjorth Parameters - _ / ,„■>,ry Mean* Complexity

[0141]

[0142] \ ‘

[0143] 103 = r}5116 Interquartile Range*lnD

[0144]

[0145] SkewnessXstsfn — 1)S3'

[0146]

[0147] 104 Kurtosis •v1— A)z4117 Mean or Median

[0148] Absolute Deviation* M 4A A nD n mia ndlfxj

[0149]

[0150] 105 Root Mean Square;n118 Coefficient of

[0151] y=Variation DK=(S / x )100

[0152]

[0153] J”1=1

[0154] 106 Shape Factor j _ 119 Standard Error S x-=S / n

[0155] SF = X^K~ 71 i)— — i n

[0156] i = l

[0157] * The feature was calculated using MATLAB; x represents the samples, n represents the number of samples, S2 represents the variance of the signal x, S12 represents the variance of the 1st derivative of the signal x, and S22 represents the variance of the 2nd derivative of the signal x.

[0158] Table 2: 25 statistical features obtained from the flow-volume curve The distributions of the features were analyzed, and it was determined that they did not exhibit a normal distribution. Therefore, non-parametric methods were preferred to calculate the correlation coefficients.

[0159] Different feature selection algorithms were tested, and it was observed that the correlation values of the features were determined more accurately with Spearman.

[0160] The Spearman correlation values of the features were calculated and ranked from high to low (Table 3).

[0161] 10 sub-data sets were created by selecting the features ranked from high to low according to correlation coefficient values in the range of 5%-50%.

[0162] The data set containing all features was also determined as the 11th data set.

[0163] Artificial intelligence-based models were created using the data sets, and it was aimed to perform OSA diagnosis with the model offering the optimal feature-performance balance.

[0164] Spearma Spearma Spearma n Spearma Spearma n n correlatio n n correlatio correlatio Featur n Featur correlatio Featur correlatio Featur n Featur n e coefficien e n e n e coefficien e coefficien Numb t Numb coefficien Numb coefficien Numb t Numb t er (P) er t (P) er t(p) er (P) er (P) 71 0,52 15 0,14 43 0,10 34 0,06 74 0,04 50 0,29 66 0,14 17 0,10 91 0,06 38 0,04 84 0,26 3 0,13 86 0,10 109 0,06 98 0,03 83 0,25 16 0,13 18 0,10 90 0,06 48 0,03 51 0,25 22 0,13 29 0,10 77 0,06 108 0,03 85 0,24 56 0,13 68 0,10 28 0,06 64 0,02 52 0,24 14 0,13 25 0,10 42 0,06 47 0,02 76 0,23 110 0,13 23 0,10 10 0,06 49 0,02 104 0,21 93 0,13 67 0,09 61 0,06 107 0,02 33 0,19 82 0,13 75 0,09 73 0,05 31 0,02 37 0,19 53 0,12 12 0,09 26 0,05 63 0,02 21 0,18 72 0,12 113 0,09 30 0,05 39 0,02 103 0,18 101 0,12 119 0,09 69 0,05 9 0,02 20 0,18 1 0,12 80 0,08 32 0,05 89 0,02 46 0,17 102 0,12 70 0,08 117 0,05 60 0,01 19 0,17 2 0,11 11 0,08 79 0,05 59 0,01 5 0,16 106 0,11 116 0,08 24 0,05 35 0,01

[0165]

[0166] 111 0,16 54 0,11 99 0,07 36 0,05 65 0,01 41 0,16 96 0,11 105 0,07 40 0,05 114 0,01 45 0,16 97 0,11 100 0,07 27 0,05 8 0,01 44 0,15 78 0,11 112 0,07 95 0,04 115 0,01 55 0,15 57 0,11 88 0,07 87 0,04 58 0,01 94 0,14 13 0,11 118 0,07 81 0,04 62 0,00 6 0,14 7 0,10 4 0,07 92 0,04

[0167]

[0168] Table 3: Spearman correlation values sorted from high to low

[0169] For classification and performance data; 55 different models performing binary classification were created using 11 different data sets and 5 different machine learning algorithms. Among the 55 models, the data percentage offering the highest performance with the minimum number of features was determined (Table 4). New models were created using feature numbers below and above the number of features corresponding to the data percentage determined according to correlation coefficients (Table 5). It was aimed to determine the optimum model offering the highest performance with the minimum features.

[0170] Features Features Method Accuracy Sensitivity Specificity Precision F- Kappa AUC (%) (%) measure

[0171] 5 71, 50, 84, 83, SVM 96,4029 0,9744 0,9508 0,9620 0,9682 0,9617 0,9807

[0172] 51, 85

[0173] 10 71, 50, 84, 83, Desicion 89,9281 0,9231 0,8689 0,9000 0,9114 0,8930 0,9682

[0174] 51, 85, 52, 76, Tree

[0175] 104, 33, 37, 21

[0176] 15 71, 50, 84, 83, NN 93,5252 0,9231 0,9508 0,9600 0,9412 0,9309 0,9790

[0177] 51, 85, 52, 76,

[0178] 104, 33, 37,

[0179] 21,

[0180] 103, 20, 46,

[0181] 19, 5, 111

[0182] 20 71, 50, 84, 83, NN 91,3669 0,9359 0,8852 0,9125 0,9241 0,9083 0,9542

[0183] 51, 85, 52, 76,

[0184] 104, 33, 37,

[0185] 21,

[0186] 103, 20, 46,

[0187] 19, 5, 111, 41,

[0188] 45, 44, 55, 94,

[0189] 6

[0190] 25 71, 50, 84, 83, Ensemble 89,9281 0,9487 0,8361 0,8810 0,9136 0,8935 0,9575

[0191] 51, 85, 52, 76, Decision

[0192] 104, 33, 37, Tree

[0193] 21,

[0194] 103, 20, 46,

[0195] 19, 5, 111, 41,

[0196] 45, 44, 55, 94,

[0197] 6,

[0198] 16, 66, 3, 16,

[0199] 22, 56

[0200] 30 71, 50, 84, 83, Desicion 88,4892 0,9103 0,8525 0,8875 0,8987 0,8777 0,9477

[0201] 51, 85, 52, 76, Tree

[0202] 104, 33, 37,

[0203]

[0204] 21, 103, 20, 46,

[0205] 19, 5, 111, 41,

[0206] 45, 44, 55, 94,

[0207] 6,

[0208] 16, 66, 3, 16,

[0209] 22, 56, 14,

[0210] 110, 93, 82,

[0211] 53, 72

[0212] 71, 50, 84, 83, Desicion 89,2086 0,9231 0,8525 0,8889 0,9057 0,8855 0,9310 51, 85, 52, 76, Tree

[0213] 104, 33, 37,

[0214] 21,

[0215] 103, 20, 46,

[0216] 19, 5, 111, 41,

[0217] 45, 44, 55, 94,

[0218] 6,

[0219] 16, 66, 3, 16,

[0220] 22, 56, 14,

[0221] 110, 93, 82,

[0222] 53, 72

[0223] 101, 1, 102, 2,

[0224] 106, 54

[0225] 71, 50, 84, 83, Desicion 85,6115 0,8846 0,8197 0,8625 0,8734 0,8471 0,9307 51, 85, 52, 76, Tree

[0226] 104, 33, 37,

[0227] 21,

[0228] 103, 20, 46,

[0229] 19, 5, 111, 41,

[0230] 45, 44, 55, 94,

[0231] 6,

[0232] 16, 66, 3, 16,

[0233] 22, 56, 14,

[0234] 110, 93, 82,

[0235] 53, 72

[0236] 101, 1, 102, 2,

[0237] 106, 54, 96,

[0238] 97, 78, 57, 13,

[0239] 7

[0240] 71, 50, 84, 83, NN 89,9281 0,8974 0,9016 0,9211 0,9091 0,8926 0,9744 51, 85, 52, 76,

[0241] 104, 33, 37,

[0242] 21,

[0243] 103, 20, 46,

[0244] 19, 5, 111, 41,

[0245] 45, 44, 55, 94,

[0246] 6,

[0247] 16, 66, 3, 16,

[0248] 22, 56, 14,

[0249] 110, 93, 82,

[0250] 53, 72

[0251] 101, 1, 102, 2,

[0252] 106, 54, 96,

[0253] 97, 78, 57, 13,

[0254] 7

[0255] 43, 17, 86, 18,

[0256] 29, 68

[0257] 71, 50, 84, 83, NN 89,9281 0,8718 0,9344 0,9444 0,9067 0,8921 0,9586 51, 85, 52, 76,

[0258] 104, 33, 37,

[0259] 21,

[0260] 103, 20, 46,

[0261] 19, 5, 111, 41,

[0262] 45, 44, 55, 94,

[0263] 6,

[0264] 16, 66, 3, 16,

[0265]

[0266] 22, 56, 14, 110, 93, 82,

[0267] 53, 72

[0268] 101, 1, 102, 2,

[0269] 106, 54, 96,

[0270] 97, 78, 57, 13,

[0271] 7

[0272] 43, 17, 86, 18,

[0273] 29, 68, 25, 23,

[0274] 67, 75, 12, 113

[0275] 100 All Desicion 86,3309 0,8846 0,8361 0,8734 0,8790 0,8546 0,9386

[0276] Tree

[0277]

[0278] Table 4: Data percentage offering the highest performance with the minimum number of features among 55 models

[0279] Feature Features Methods Accuracy Sensitivity Specificity Precision F- Kappa AUC Quantit (%) measure

[0280] y

[0281] 1 71 kNN 76,9784 0,8974 0,6066 0,7447 0,8140 0,7590 0,770

[0282] 1 2 71, 50 NN 76,9784 0,8333 0,6885 0,7738 0,8025 0,7565 0,826

[0283] 2 3 71, 50, 84 SVM 93,5252 0,9487 0,9180 0,9367 0,9427 0,9311 0,957

[0284] 3 4 71, 50, 84, 83 NN 96,4029 0,9615 0,9672 0,9740 0,9677 0,9617 0,982

[0285] 7 5 71, 50, 84, 83, NN 97,1223 0,9744 0,9672 0,9744 0,9744 0,9694 0,988 51 4 6 71, 50, 84, 83, SVM 96,4029 0,9744 0,9508 0,9620 0,9682 0,9617 0,980 51, 85 5 7 71, 50, 84, 83, Desicion 94,2446 0,9359 0,9508 0,9605 0,9481 0,9386 0,970 51, 85, 52 Tree 2 8 71, 50, 84, 83, NN 92,8058 0,9103 0,9508 0,9595 0,9342 0,9231 0,972

[0286]

[0287] 51, 85, 52, 76 9 Table 5: New models created using feature numbers below and above the number of features corresponding to the data percentage determined according to correlation coefficients

[0288] Decision Tree (DT), Support Vector Machine (SVM), k-nearest neighbor (kNN), Ensemble Decision Tree (ET), and Neural Network (NN) were used as machine learning methods. Spirometers most commonly used in the clinical environment are connected to computers having low / medium level processor capacities suitable for daily use. Models to be added to these types of devices must be suitable for working on these devices. The machine learning algorithms used were preferred due to their not needing computers having high processor capacity, their applicability to embedded systems, their training times being short, and their ability to perform high-performance classification. Each data set was divided as 70% training - 30% test using the systematic sampling method. The models were trained by performing 5-fold cross-validation and hyperparameter optimization. The trained models were tested using unseen datasets. Test results were evaluated using 7 performance evaluation criteria. In order to calculate the performance parameters of the models, a confusion matrix was created first. Accuracy, sensitivity, specificity, precision, F-measure, kappa value, and AUC were calculated using the confusion matrix.

[0289] Spearman correlation grading:

[0290] Grade Correlation

[0291] ρ = 0 None

[0292] 0 < |p| < 0.19 Very Low

[0293] 0.20 ≤ |ρ| ≤ 0.39 Low

[0294] 0.40 ≤ |ρ| ≤0.59 Moderate

[0295] 0.60 ≤ |ρ| ≤ 0.79 High

[0296] 0.80 ≤ |ρ| ≤ 1 Very High

[0297] 1.00 Perfect

[0298]

[0299] Accuracy gives information about the ratio of correctly predicted data and is expressed as a percentage. The larger the accuracy, the better it is. It does not give information about the correct prediction ratio of patients or healthy individuals. Therefore, it is not a sufficient criterion alone.

[0300] Sensitivity and Specificity are the rates of predicting patients and healthy individuals, respectively. They vary between 0-1. The closer they are to 1, the better it is.

[0301] Precision expresses how many of those predicted as patients are patients.

[0302] F-measure gives the harmonic mean of sensitivity and precision. It varies between 0 and 1; the closer it is to 1, the better it is.

[0303] Kappa indicates the agreement between two observers in the evaluation of categorical variables. It varies between +1 and -1. +1 means a perfect agreement, while -1 means a perfect disagreement. Kappa values above 0.81 are considered quite good.

[0304] The ROC curve is the “(1 -specificity) - sensitivity” graph. AUC represents the area under this graph. It determines the power of discrimination between classes. It is expressed between 0-1. A value close to 1 indicates a good discrimination.

Claims

1. CLAIMS1. A method for processing data directed towards the diagnosis of obstructive sleep apnea (OSA) according to characteristic and statistical features of value curves calculated values obtained based on breath speed, intensity, and frequency by performing respiratory function tests on many selected patients, characterized in that it comprises;3.determining the start points of flow-volume curve maneuvers according to breathing in and out data,4.obtaining separate flow-volume curves using the start points,5.calculating the total lung capacity (TLC) value from the difference between the start and end volume values of each flow-volume curve,6.selecting the flow-volume curve having the highest TLC value to be processed in signal processing stages,7.extracting characteristic and statistical features from the selected curve, creating a database with the extracted features,8.selecting non-parametric methods via distribution analysis of the features and calculating their impact levels in determining the class label,9.creating different models using machine learning methods and feature selection, evaluating the performances of the models with certain criteria, determining whether the individual belongs to the obstructive sleep apnea class as a result of the evaluation.

2. The diagnostic method directed towards obstructive sleep apnea according to claim 1, characterized in that model performances are evaluated using evaluation criteria of accuracy, sensitivity, specificity, precision, F-measure, kappa value, and area under the receiver operating characteristics curve (AUC).

Images