Mobile QR Code QR CODE

  1. (Department of Computer Software, ICT, University of Science and Technology, Daejeon 34113, Korea {synam0216, johnlorenzobautista} )
  2. (Emotion(Brain-Emotion) Research Laboratory, Electronics and Telecommunications Research Institute, Daejeon 34129, Korea, )

Apnea, Sleep apnea, Sleep disorder breathing, Apnea-hypopnea index (AHI), Respiratory disturbance index (RDI), Obstructive sleep apnea, Respiration-induced intensity variation (RIIV), Respiratory effort strength index (RESI), Photoplethysmogram, Wearable device

1. Introduction

Sleep apnea and sleep fragmentation are sleeping disorders characterized by the partial or complete cessation of breathing occurring many times during sleep, resulting in reduced (hypopnea) or absent (apnea) airflow lasting for at least 10 seconds and associated with either cortical arousal or a fall in blood oxygen saturation [1]. Frequent sleep disturbances may cause high blood pressure, type 2 diabetes, atrial fibrillation, heart failure, coronary artery disease, stroke, and death [2]. In addition, sleep apnea increases comorbid health risks, such as insomnia and depression [3-5]. The diagnosis of sleep apnea can be achieved through a number of different tests and techniques. The breathing activity of sleep apnea patients must be examined directly using an expensive and time-consuming procedure called polysomnography (PSG), which measures several physiological parameters, such as nasal-airflow, blood pressure, electrocardiogram, electroencephalogram, electromyogram, and blood oxygen saturation during sleep. Sleep apnea can be diagnosed with either home- or laboratory-based sleep testing, and effective treatments are available using simple wearable devices [6,7].

According to an estimation of the global prevalence and burden of obstructive sleep apnea, reliable prevalence data for obstructive sleep apnea were available for 16 countries from 17 studies. Using the AASM 2012 diagnostic criteria and AHI threshold values of five or more events per hour and 15 or more events per hour, it was estimated that 936 million (95% CI 903-970) adults aged 30-69 years (men and women) have mild to severe obstructive sleep apnea and 425 million (399-450) adults aged 30-69 years have moderate to severe obstructive sleep apnea globally. The number of affected individuals was highest in China, followed by the USA, Brazil, and India [8]. The definition of sleep apnea is the cessation of oronasal flow for more than 10 seconds. The condition can be diagnosed using the apnea-hypopnea index (AHI) by reliably classifying the severity and the frequency of apneic events per hour during sleep. Normal individuals have an AHI of less than 5, but severe sleep apnea patients have greater than 30 events per hour [9]. The respiratory disturbance index (RDI) is also used for reporting respiratory events during sleep. The RDI is the average number of respiratory event-related arousals (apnea and hypopnea) per hour of sleep [10].

Because the goal of ventilation, which is the movement of air into and out of the lungs, is to maintain the homeostasis of the arterial oxygen and carbon dioxide concentrations, the recognition of normal respiration and apnea is of great importance for surgical and emergency patients requiring immediate medical care. Many studies have examined the relationship between respiration and blood pressure in respiration physiology from the 1900s. They show that respiration and blood pressure have a strong correlation [11,12]. Several frequency components of photoplethysmogram (PPG) caused by breathing were presented [12]. A respiratory signal can be measured via the nasal airflow and abdomen/thoracic movement. These measurement techniques are inconvenient and uncomfortable for use in daily life. Several studies have reported that respiration influences electrocardiogram (ECG) and PPG signals. A new method for extracting the respiratory signal using PPG, called the respiration-induced intensity variations (RIIV), has been suggested. The RIIV is well synchronized with the breathing rhythm [12], and can be extracted from a PPG signal, which is a non-invasive, straightforward, and convenient technique that measures the light absorption of the skin blood vessels constantly undergoing spontaneous variations in volume and concentration [13]. Respiratory-induced variations of the blood cause blood volume variations on both the arterial and venous sides. This makes a change in the baseline PPG signal with a frequency of 0.2-0.45 Hz as a normal adult respiratory period [14]. Several methods for extracting the RIIV signal from the PPG signal are presented. Although it is difficult to select the precise respiration frequency without considering various conditions, a bandpass filter can extract the RIIV signal [12,15]. Power spectrum estimation is presented using a pulse width variability (PWV), pulse amplitude variability (PAV), and pulse rate variability (PRV) of the PPG signal [16-18]. RIIV signal extraction using both accurate non-parametric methods (e.g., time-frequency spectral methods) [19,20], and parametric methods (e.g., probability autoregressive model) were proposed [11]. Although several methods to extract an RIIV signal have been presented, there is currently no method to recognize the normal respiration and apnea to indicate the severity of sleep apnea using an RIIV signal. Therefore, this paper proposes a respiration effort strength index (RESI), measuring the degree of respiration effort to recognize normal or apneic respiration using the frequency domain analysis of a RIIV signal by comparing the lack of a frequency component in apneic respiration against normal respiration in real-time. RESI, unlike PSG, requires only a PPG signal without needing to monitor various physiological signals, such as nasal airflow, electroencephalography (EEG), electrooculography (EOG), electromyography (EMG), and ECG. RESI can ideally classify the apnea types as central and obstructed.

2. Materials and Methods

2.1 Data Acquisition and Processing

The physiological signals, PPG, and referenced thoracic respiration movement signal were collected to recognize apnea using an RIIV signal. The measurement system consisted of two parts: (1) Direct measurement of the respiratory signals of thoracic expansion and contraction using RSP100C from Biopac Systems Inc., and (2) Measured PPG signals using the developed PPG detection device, as shown in Fig. 2. The RSP100C is operated simultaneously to verify the correct respiratory and apneic episode with a PPG sensing device. A prototype wristwatch-type wearable device that includes a reflective optical sensor with a green LED (535 nm) located on the bottom of the device was used to measure the PPG signal on top of the wrist. An integrated analog front-end heart rate monitor device, which senses the PPG signal, and a microprocessor (Cortex-M0, 32 bit), which calculates the RIIV signal from a PPG signal, were embedded, as shown in Fig. 2. The sampling frequency of the RSP100C was 100 Hz, while the PPG sensing device was 64 Hz.

Data acquisition was carried out on five healthy male volunteers aged 30 to 55 years of age (median: 43.8 years old) after obtaining informed consent. During a 12-minute session, volunteers ceased their respiration twice for about a minute during the 5$^{\mathrm{th}}$ and 9$^{\mathrm{th}}$ minutes. In this paper, the respiratory signal from the PPG was extracted using a digital filtering technique [12,15]. The RIIV signal was extracted from the second-order infinite impulse response (IIR) Bessel bandpass filter and third-order IIR elliptical lowpass filter. The elliptic low pass filter was used to eliminate the high-frequency components of the RIIV signal after the Bessel bandpass filter. Table 1 lists the filter cutoff frequency. The extracted RIIV signal, PPG, thoracic respiratory movement, and RIIV signal during normal respiration signals were synchronized in time to allow easy comparison, as shown in Fig. 3.

Fig. 1. Measurement system: (a) RSP100C and wristwatch type PPG-sensing device with a USB connector (each device has a wired connection with a PC); (b) Flow diagram of sensing PPG signal and respiration movement and calculating RIIV.
Fig. 2. PPG sensing device: (a) Reflected optical sensor located on the back of a wristwatch type sensing device; (b) Prototype PPG sensing wearable device tested during experiments.
Fig. 3. Comparison of the respiratory signal, RIIV, (black) from PPG (blue) with the directly measured respiration movement signal (red) during normal respiration.
Table 1. Filter specifications.

Cutoff frequency (Hz)

Filter order


1.0 ~ 3.0


Bessl BPF (IIR)

0.15 ~ 0.38


Elleptic LPF (IIR)



2.2 Respiratory Signal Analysis

To quantify the RESI from the RIIV signal, the PPG, including both the normal respiration and apnea (voluntary cessation of respiration are measured, every 5$^{\mathrm{th}}$ and 9$^{\mathrm{th}}$ minute for at least 30 seconds) for 12 minutes, as shown in Fig. 4. This directly measures the respiration movement of the thoracic area, which can be verified with the exact apneic time, as shown in Fig. 4(a). The periodicity of three times - the normal respiration, first apnea, and second - can be differentiated and identified in Figs. 4(a) and (b). Power spectral analysis of respiration and apnea in Fig. 4(b) was performed using Welch’s power spectral density estimation method [21]. Fig. 5 shows the detailed segmentation and overlap analysis of Welch’s power spectral density. The parameters of 24 sample overlaps, 512 window size, and 512- point radix-2 Cooley-Tukey FFT algorithm with a Hanning window using a 1,000 points (15.625 seconds) of respiratory signal (12 minutes, 46,080 samples) were utilized. Respiration analysis using the power spectral is the power spectral ratio of the normal respiration frequency band to the reference frequency band. The normal respiration rate of an adult is baselined at 12 times/min (0.1-0.3 Hz), and a reference frequency of under 0.2 Hz was used to remove various skin conditions. The calculated results of Welch’s power spectral density estimation of normal respiration of arbitrary time, first apnea and second apnea in Fig. 4(b) are illustrated in Fig. 6(a). Fig. 6 presents the power spectral density of five volunteers. The power spectral density of normal respiration (blue) surpassed the first apnea and second apnea (red) at the frequency 0.1-0.3 Hz. The result of power spectral density of the first apnea was compared with the arbitrary normal respiration before the first apnea. The second apnea was compared with the arbitrary normal respiration between the first and second apnea for each volunteer, as shown in the left and right of Fig. 6, respectively.

Fig. 4. Respiratory and RIIV signal from PPG. The normal respiration and two apnea events can be seen (first apnea : 5 minute, second apnea : 9 minute for approximately 30 seconds): (a) directly measured respiration movement signal; (b) RIIV signal from PPG.
Fig. 5. Analysis of the respiratory signal by Welch’s power spectral density.
Fig. 6. Power spectral density of normal respiration (blue) and apnea (red) of five volunteers. Left figure: normal respiration of arbitrary time and first apnea, and right figure: normal respiration of arbitrary time and apnea.

3. Results

The respiratory signal and respiratory effort of normal respiration were very periodic, as shown in Fig. 4. However, there was no thoracic respiratory movement in Fig. 4(a), and the fluctuation in the venous and arterial blood caused by respiration was aperiodic, as shown in Fig. 4(b) during the voluntary cessation of respiration, namely the first apnea and second apnea. The respiratory muscle called the diaphragm, has a range of movement of approximately 1.5 cm for shallow respiration. On the other hand, the diaphragm has a range of movement of 7-13 cm for deep breathing [21]. For respiratory homeostasis, the autonomic respiratory center in the medulla oblongata controls the timing and depth of respiration to the apneic episode and normal respiration in Fig. 4(b). In effect, the autonomic function of respiration struggles randomly to breathe more frequently during voluntary cessation than normal respiration because of respiratory homeostasis. This circumstance clearly shows the respiration effort [21]. Therefore, the respiratory effort strength index (RESI) can be defined in the index of striving to breathe during normal respiration and apnea using a deficiency of power spectral density at the frequency of normal respiration.

3.1 RESI

The RESI was calculated using (1). X$_{PS}$$\textit{(f)}$ is the power spectrum of a respiratory signal in Fig. 4(b) at a frequency of f.

$ RESI=\frac{X_{PS}\left(0.1hz<f<0.3hz\right)}{X_{PS}\left(0hz<f<0.5hz\right)} $

$X_{PS}\left(0.1hz<f<0.3hz\right)$ and $X_{PS}\left(0hz<f<0.5hz\right)$ mean the total spectral density under 0.5 Hz as a reference spectral density, which can be used for various skin conditions on top of the wrist, such as the depth of the dermis and distribution of blood vessels (2), and the spectral density between $\textit{0.1hz}$ to $\textit{0.3hz}$ as a frequency of normal respiration (3), respectively.

$ X_{PS}\left(0hz<f<0.5hz\right)=\sum _{f=0hz}^{0.5hz}\left| \hat{x}\left(f\right)\right| ^{2} \\ $
$ X_{PS}\left(0hz<f<0.3hz\right)=\sum _{f=0.1hz}^{0.3hz}\left| \hat{x}\left(f\right)\right| ^{2} $

Where XPS(f) is obtained as:

$ X_{PS}\left(f\right)=~ \left| X\left(f\right)\right| ^{2}=~ X_{real}\left(x\right)^{2}+~ X_{imag}\left(x\right)^{2} $

The $\textit{X(f)}$ is composed of real, X$_{real}$$\textit{(f)}$, and imaginary, X$_{imag}$$\textit{(f)}$ components.

4. Discussion

Fig. 7 shows the RESI calculated from (1) and filtered RESI of five volunteers to recognize sleep apnea. The RESI was filtered using a 2$^{\mathrm{nd}}$ order Bessel low pass filter on MATLAB by MathWorks Inc., with a cutoff frequency = 0.1 Hz. RESI was used to recognize apnea. The results of filtered RESIs for each apneas were relatively lower at the 5$^{\mathrm{th}}$ and 9$^{\mathrm{th}}$ minute of a 12-minute interval than with normal respiration. The filtered RESIs were scored between 0.5 to 0.7 for normal respiration and well-matched in time with normal respiration and two apneas of each volunteer.

The concurrence of filtered RESI and thoracic respiration movement was measured directly, as shown in Fig. 8. The filtered RESI starts with a rapid decline at the beginning of the apnea. Even if the volunteer stops respiration by closing their nose for more than 30 seconds, respiration movement of the diaphragm can still be observed, as shown in Fig. 8(c). The results suggest that the respiratory system maintains and persists a respiration effort caused by the autonomic nervous system (ANS) even during a simulated apneic event. The filtered RESI of the first apnea is more symmetrical at the lowest filtered RESI than the second apnea in Fig. 8.

Fig. 8 shows the filtered RESI of five volunteers on the same graph, which were scaled up and down to avoid overlap for better visual comparison. Before the first apnea, the first normal respiration is quieter than the second and third normal respiration. After the second apnea, the RESI, there are more fluctuations compared with the values before the first apnea. The filtered RESI of the first apnea is easy to recognize compared to the filtered RESI of the second apnea. This suggests that the autonomic nervous system, particularly the respiratory control system, aims to maintain adequate gas levels of CO$_{2}$, H$^{+}$, and pH within the blood vessels of a normal state [23,24].

Fig. 7. RESI calculated from (1) shown in left, and filtered RESI using 2nd order Bessel low pass filter shown in right (a)~(e) respectively.
Fig. 8. Synchronized view of the filtered RESI (red) with thoracic respiration movement (blue). Filtered RESI for each volunteer are shown in Fig. 9 as (a) black dotted; (b) blue solid; (c) red dotted; (d) black solid; (e) red solid lines.
Fig. 9. The filtered RESI of the five volunteers: The first and second apnea episodes are highlighted by the transparent gray boxes.

5. Conclusion

A method was introduced to recognize sleep apnea using a PPG sensor equipped on the back of a wristwatch-type wearable device. An IIR Bessel digital BPF filter with 0.15-0.38 hz was used to extract blood volume variation in arterial and venous blood vessels. The respiration effort strength index was calculated from the variations of blood volume caused by respiration using Welch’s power spectral density estimation method.

The respiration effort strength index reflects the blood pressure and is strongly correlated with respiration. The RESI is the score of striving to breathe for both normal and apneic respiration. RESI can distinguish the lack of frequency component compared with normal respiration during breathing. In this paper, RESI is used to recognize the apneic episode with two parts: (1) sensing and extracting the RIIV using a wristwatch type wearable device, and (2) analyzing the power spectral density of RIIV using a personal computer. RESI can be used to diagnose sleep apnea for easy-to-use self-diagnosis devices.

If there is no respiration effort in cases, such as central apnea, the RESI may be much lower than an obstructed apnea. As a result, RESI can also be used to differentiate between obstructed sleep apnea and central apnea. Nevertheless, further study will be needed.


This work was supported by the Korea Evaluation Institute of Industrial Technology (KEIT) grant funded by the Korea government (MOTIE) (No. 20012603, Development of Emotional Cognitive and Sympathetic AI Service Technology for Remote (Non-face-to-face) Learning and Industrial Sites).


Gottlieb Daniel J., Naresh M. Punjabi., 2020, Diagnosis and management of obstructive sleep apnea: a review., Jama, Vol. 323, No. 14, pp. 1389-1400DOI
Dempsey Jerome A., et al. , 2010, Pathophysiology of sleep apnea., Physiological reviews, Vol. 90, No. 1, pp. 47-112DOI
Frost and Sullivan , 2016, Hidden Health Crisis Costing America Billions - Underdiagnosing and Undertreating Obstructive Sleep Apnea Draining Healthcare System, American Academy of Sleep MedicineURL
Luyster Faith S., Daniel J. Buysse., Patrick J. Strollo Jr., 2010, Comorbid insomnia and obstructive sleep apnea: challenges for clinical practice and research., Journal of Clinical Sleep Medicine, Vol. 6, No. 2, pp. 196-204DOI
Barker Sarah J., 2011, Obstructive sleep apnoea and type 2 diabetes: A wake-up call., Journal of diabetes nursing, Vol. 15, No. 10DOI
Manoni Alessandro, et al. , 2020, A new wearable system for home sleep apnea testing, screening, and classification., Sensors, Vol. 20, No. 24, pp. 7014DOI
Hilmisson Hugi, Stephen Berman., Solveig Magnusdottir., 2020, Sleep apnea diagnosis in children using software-generated apnea-hypopnea index (AHI) derived from data recorded with a single photoplethysmogram sensor (PPG)., Sleep and Breathing, Vol. 24, No. 4, pp. 1739-1749DOI
enjafield Adam V., et al. , 2019, Estimation of the global prevalence and burden of obstructive sleep apnoea: a literature-based analysis., The Lancet Respiratory Medicine, Vol. 7, No. 8, pp. 687-698DOI
Collop Nancy A., et al. , 2011, Obstructive sleep apnea devices for out-of-center (OOC) testing: technology evaluation., Journal of Clinical Sleep Medicine, Vol. 7, No. 5, pp. 531-548DOI
Peled Nir, et al. , 1998, Contributions of hypoxia and respiratory disturbance index to sympathetic activation and blood pressure in obstructive sleep apnea syndrome., American journal of hypertension, Vol. 11, No. 11, pp. 1284-1289DOI
Lewis Thomas., 1908, Studies of the relationship between respiration and blood-pressure: Part II. Facts bearing on the relationship of different factors in the production of respiratory curves of blood-pressure1., The Journal of physiology, Vol. 37, No. 3, pp. 233DOI
A. C. DORNHORST , P. HOWARD , G. L. LEATHART , 1952, Respiratory variations in blood pressure., Circulation, Vol. 6, No. 4, pp. 553-558DOI
Nilsson Lena M., 2013, Respiration signals from photoplethysmography., Anesthesia & Analgesia, Vol. 117, No. 4, pp. 859-865DOI
Lindberg L-G. , H. Ugnell. , P. Å. Öberg. , 1992, Monitoring of respiratory and heart rates using a fibre-optic sensor., Medical and Biological Engineering and Computing, Vol. 30, No. 5, pp. 533-537DOI
Rybynok V. O. , P. A. Kyriacou. , 2010, Beer-lambert law along non-linear mean light pathways for the rational analysis of photoplethysmography., Journal of Physics: Conference Series, IOP Publishing, Vol. 238, No. 1DOI
Meredith David J., et al. , 2012, Photoplethysmographic derivation of respiratory rate: a review of relevant physiology., Journal of medical engineering & technology, Vol. 36, No. 1, pp. 1-7DOI
Nakajima K., et al. , 1993, Photoplethysmographic measurement of heart and respiratory rates using digital filters., Proceedings of the 15th Annual International Conference of the IEEE Engineering in Medicine and Biology Societ. IEEEDOI
Lázaro Jesús, et al. , 2011, Deriving respiration from the pulse photoplethysmographic signal., 2011 Computing in Cardiology. IEEEURL
Karmakar , Chandan , et al. , 2013, Detection of respiratory arousals using photoplethysmography (PPG) signal in sleep apnea patients., IEEE journal of biomedical and health informatics, Vol. 18, No. 3, pp. 1065-1073DOI
Leier, Mairo , Gert Jervan. , Wilhelm Stork. , 2014, Respiration signal extraction from photoplethysmogram using pulse wave amplitude variation., 2014 IEEE International Conference on Communications (ICC). IEEEDOI
Ellis H., 2012, The respiratory system., Anaesthesis and Intensive Care Medicine, Vol. 12, pp. 533-538DOI
Welch Peter., 1967, The use of fast Fourier transform for the estimation of power spectra: a method based on time averaging over short, modified periodograms., IEEE Transactions on audio and electroacoustics, Vol. 15, No. 2, pp. 70-73DOI
Guz A., 1997, Brain, breathing and breathlessness., Respiration physiology, Vol. 109, No. 3, pp. 197-204DOI
Flinta Irena. , Piotr Ponikowski. , 2016, Relationship between central sleep apnea and Cheyne− Stokes respiration., International journal of cardiology 206, pp. s8-S12DOI


Seungyoon Nam

Seungyoon Nam joined the Electronics and Telecommunications Research Institute (ETRI) in 2013, where he has been involved in developing the emotion recognition Research Section of ETRI. He is currently a Ph.D. Student at University of Science and Technology since 2016. His research interests include sensing data on the wearable device and bio-signal analysis for human care.

John Lorenzo Bautista

John Lorenzo Bautista is currently attending the University of Science and Technology (UST) in Daejeon, South Korea as a Ph.D. student majoring in Computer Software since 2017. He graduated Master in Engineering in Computer Engineering from Hanbat National University, and his Bachelor of Science in Computer Science from La Consolacion College Manila. He joined Electronics and Telecommunications Research Institute (ETRI) in 2017 as part of his Ph.D. program and is involved with projects from the Emotion Recognition Research Section of ETRI. His research interests include Machine Learning Techniques relating to Speech Recognition Technology, Natural Language Processing, and Biosignals Processing.

Chanyoung Hahm

Chanyoung Hahm received the MS degree in the School of Information and Communications from Gwangju Institute of Science and Technology (GIST), Republic of Korea, in 2002. He joined Electronics and Tele-communications Research Institute (ETRI) in 2013, where he has been involved in the development of the emotion recognition Research Section of ETRI. His research interests include wearable devices, emotion recognition, and bio-signal analysis for human care.

Hyunsoon Shin

Hyunsoon Shin received her Ph. D. degree in computer science. Currently, she is a project leader in the brain-emotional research division of the Electronics and Telecommunications Research Institute (ETRI), a professor at the University of Science and Technology (UST), and president of the Korea Emotion Information and Communication Technology Industry Association (EICT). Her research interests are in the areas of emotion recognition algorithms, bio-signal & emotion signal processing, wearable devices, bio-signal analysis for human care, physiology & voice(speech) emotion recognition algorithms, HX (Human eX) processing technology, big-data processing, data mining, deep learning, deep analysis, emotion data knowledge, and protocol for emotional signal & information.