A Vision-Based Respiration Monitoring System for ... - IEEE Xplore

1904

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 63, NO. 9, SEPTEMBER 2016

A Vision-Based Respiration Monitoring System for Passive Airway Resistance Estimation Sarah Ostadabbas∗ , Member, IEEE, Nordine Sebkhi, Student Member, IEEE, Mingxi Zhang, Salman Rahim, Larry J. Anderson, Frances Eun-Hyung Lee, and Maysam Ghovanloo, Senior Member, IEEE

Abstract— Objective: Airway resistance is the mechanical cause of most of the symptoms in obstructive pulmonary disease, and can be considered as the primary measure of disease severity. A low-cost and noninvasive method to measure the airway resistance that does not require patient effort could be of great benefit in evaluating the severity of lung diseases, especially in patient population that are unable to use spirometry, such as young children. Methods: The Vision-Based Passive Airway Resistance Estimation (VB-PARE) technology is a passive method to measure airway resistance noninvasively. The airway resistance is estimated from: 1) airflow extracted from processing depth data captured by a Microsoft Kinect, and 2) Pulsus Paradoxus extracted from a pulse oximeter (SpO2 ). Results: To verify the validity and accuracy of the VB-PARE, two phases of experiment were conducted. In Phase I, spontaneous breathing data was collected from 14 healthy participants with externally induced airway obstruction, and the accuracy of 76.2± 13.8% was achieved in predicting three levels of obstruction severity. In Phase II, VB-PARE outputs were compared with the clinical results from 14 patients. VB-PARE estimated the tidal volume with an average error of 0.07±0.06 liter. Also, patients with airway obstruction were detected with 80% accuracy. Conclusion: Using the information extracted from Kinect and SpO2 , here, we present a quantitative method to measure the severity of airway obstruction without requiring active patient involvement. Significance: The proposed VB-PARE system contributes to the state-of-art respiration monitoring methods by expanding the idea of passive and noninvasive airway resistance measurement. Index Terms—Airway resistance, microsoft kinect, obstructive pulmonary diseases, photoplethysmography (PPG), pulmonary function tests (PFTs), pulsus paradoxus, respiration monitoring.

I. INTRODUCTION ITH active obstructive pulmonary diseases, airway obstruction increases and the degree of increase indicates the extent of altered lung function and severity of the disease. Clinicians managing these illnesses need to determine the severity of airway obstruction in order to efficiently and safely manage the patients suffering from these complications.

W

Manuscript received June 26, 2015; revised October 06, 2015; accepted November 27, 2015. Date of publication December 04, 2015; date of current version August 18, 2016. Asterisk indicates corresponding author. ∗ S. Ostadabbas is with the Augmented Cognition Lab, Department of Electrical and Computer Engineering, Northeastern University, Boston, MA 02115 USA (e-mail: [email protected]). N. Sebkhi, M. Zhang, S. Rahim, and M. Ghovanloo are with the GT-Bionics Lab, School of Electrical and Computer Engineering, Georgia Institute of Technology. L. J. Anderson is with the Department of Pediatrics, Emory Hospital, Children’s Center. F. E.-H. Lee is with the Division of Pulmonary, Allergy, and Critical Care Medicine, Emory University. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TBME.2015.2505732

In this paper, we consider airway resistance, R, defined in (1) as a direct measure of the airway obstruction R=

Pair − Plung ΔP = Q Q

(1)

where Q is the airflow, Plung is the pressure inside the lung (alveolar pressure), and Pair is the atmospheric pressure [1]. Elevation in airway resistance increases the differences in lung pressures required for effective inspiration and expiration and decreases airflow for given differences in pressure. Therefore, simultaneous measurement of airflow and lung pressure is required to accurately quantify the airway resistance. Based on the method initially developed in [2], we use the airflow detection algorithm presented in Section III-B to estimate Q from changes in the chest volume using a commercially available infrared depth-sensor, the Microsoft Kinect. As for Plung , directly measuring alveolar pressure requires an invasive procedure by placing pressure sensor deep in the lung [3]. We have developed a surrogate model using Photoplethysmography (PPG) data to substitute the numerator of (1) with a measure that correlates with the changes in intrathoracic pressure during breathing, and corresponds to the force required for breathing, also called breathing effort, as described in Section IV. In this paper, using the information obtained from Kinect (airflow) and PPG (breathing effort), we present a method to provide a quantitative and reliable measure of lung function and severity of airway obstruction without requiring active patient involvement, unlike spirometry [4]. We have devised a VisionBased Passive Airway Resistance Estimation (VB-PARE) system, based on this method. With every parameter in (1) being passively measured without depending on the patients cooperation or compliance, VB-PARE can be of great benefit in evaluating the severity of lung disease or respiratory function in patients who cannot comply with breathing instructions, such as pediatric patients or those who are very weak or unconscious. It should be noted that the VB-PARE does not provide an absolute measure of the airway resistance. Absolute resistance should be calibrated with a pulmonary plethysmograph, and an internally placed pressure sensor. However, changes in the chest diameter and breathing effort would be sufficient to show relative changes in the airway resistance as an estimate of airway obstruction. The rest of the paper is organized as follows. In Section II, we review the methods conventionally used as pulmonary function tests (PFTs) in comparison with the VB-PARE. Section III, presents the algorithm developed for extracting breathing pattern and airflow from the Kinect depth data. The method for

0018-9294 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

OSTADABBAS∗ et al.: VISION-BASED RESPIRATION MONITORING SYSTEM FOR PASSIVE AIRWAY RESISTANCE ESTIMATION

Fig. 1.

1905

Automatic chest bounding box, Bch e st detection algorithm using depth data from Kinect.

breathing effort estimation extracted from PPG data is explained in Section IV. Our two-phase validation procedure is described in Section V, followed by experimental results in Section VI. In Section VII, we go over the practical aspects and limitations of using VB-PARE system at home or medical clinics. Section VIII includes concluding remarks and discusses possible future work. II. RELATED METHODS AND MOTIVATION Presently, clinicians often assess the severity of altered lung function, caused by obstructive pulmonary diseases, with imprecise but easy to administer measures and signs, such as respiratory rate, chest retractions, breath sounds, and level of O2 saturation, which in turn determine the need for hospital admission and ventilatory support [5]. Of these measures, only respiration rate and O2 saturation provide quantitative measures of illness. However, it has been shown that both may be within normal range despite presence of substantial lung disease [6]. Clinicians can also use rate of airflow during maximal expiratory effort (after maximal inhalation) obtained via spirometry as a quantitative measure of obstructive airway diseases. This test, however, requires active patient cooperation and ability to inhale and exhale deeply on command and, thus, cannot be used in infants, young children, or severely ill patients. Consequently, the need for hospitalization in infant is subjectively determined by listening to the wheezing sound during chest auscultation. Moreover, peak expiratory flow, measured by peak flowmeter to assess response to a treatment, has been recognized to produce inconsistent and unreliable results [7]. In recent years, whole-body plethysmography or “body box” has been widely used for passively measuring lung volume and pressure in pulmonary clinics. Body box has sealed and rigid walls, and is equipped with instruments that allow clinicians to assess functional lung capacities and airway resistance [8]. However, its use in very young patients has been limited to research purposes. It can also be problematic for patients with claustrophobia [8]. In parallel, advanced medical imaging technologies, such as positron emission tomography/computed tomography scanners are used as passive respiratory motion tracking systems, and can provide accurate lung capacity and breathing patterns [9]. However, because of the cost and radiation exposure, these methods are not recommended for routine clinical assessment, and their use for patients with chronic obstructive pulmonary diseases is limited to the investigation of coexisting bronchiectasis [6].

A method to accurately evaluate the respiratory distress that is noninvasive, inexpensive, and does not require patient cooperation could be of substantial value for evaluating the severity of lung disease in patient populations that are unable to use spirometry, both for clinicians managing these patients and researchers studying the effect of vaccines and discovering new medications and treatment methods. In the last few years, several groups have started using the Kinect technology for noncontact respiratory monitoring, both for breathing rate detection and airflow calculation [2], [10]– [12]. Low peak airflow is consistent with the obstructive pulmonary diseases, but has poor specificity because it can be altered by applying different breathing efforts. As a result, measuring airflow by itself is not a reliable estimation of the airway resistance. Breathing effort can be derived from PPG, one of the signal produced by the pulse oximetry. This method, however, does not measure the airflow, and consequently may falsely indicate presence or absence of lung disease [13]. The proposed VB-PARE system contributes to the state-of-art respiration monitoring methods by expanding the idea of passive airway resistance measurement introduced in [2], through combining pulse oximetry and infrared depth data recordings in order to simultaneously measure both airflow and breathing effort. Airflow, Q, is indicated by changes in the chest-wall position captured by Kinect when the patient is stationary, as presented in Section III, while breathing effort is estimated from PPG as explained in Section IV. This noninvasive and easy to use technology can passively estimate airway obstruction, and may ultimately be used to quantify the severity of obstructive pulmonary diseases at hospital, physicians office, or even home environments.

III. AIRFLOW ESTIMATION USING KINECT A. Automatic Detection of the Chest Bounding Box In order to extract the breathing pattern, we focus on the patient’s chest-wall movements captured by the Kinect camera in each breathing cycle. To maximize the signal to noise ratio, the depth measurement is limited to the human body areas that optimally demonstrate the lung volume changes during breathing, i.e., the thoracic and upper abdominal areas. Fig. 1 shows the algorithm that can automatically detects the region of interest, called chest bounding box, when the patient is in the sitting position.

1906


First, we average the original depth data, ID , from the entire Kinect depth sensor field of view over a period that subject can effectively remain motionless (1 min in this case) and compute the mean image depth, IM . Then, the binary image, IB is extracted by applying a threshold window. The upper and lower thresholds are determined from the histogram of IM , defined with three bins. The lower threshold, TL , is equal to the center of the first bin, which eliminates very close pixels, such as those from the knees. The upper threshold, TU , is set to the center of the third bin and eliminates very far pixels, such as the wall in the background. Applying two consecutive connected component search on the row and column of the binary image data, and finding the connected components larger than Mx and My , produces IC X and IC Y , respectively. Mx and My are the median lengths of the row and column connected components, respectively. In the final step, we use a few tailored measurements from the subject’s upper body to anatomically adjust IC Y to the subject’s exact chest area. The anatomical adjustment is done by cropping the bottom rows of IC Y to match with the length of the subject’s torso. This step results in the final chest bounding box, Bchest . B. Airflow Calculation To estimate the respiratory airflow during spontaneous breathing, we used an enhanced version of the method previously presented in [2]. Kinect can accurately capture the rise and fall of the chest and upper abdomen wall during each breathing cycle, and lung volume changes over time is related to the chest-wall movements, as long as the subject remains stationary [14]. The chest volume, Vchest (t), at time t can be estimated by numerically integrating the depth value, dij (t), of every pixel in Bchest Vchest (t) = V0 −

dij (t)

(2)

i,j ∈Bc h e s t

where V0 is the volume between the subject’s back and the camera. Airflow is based on the changes in lung volume, Vlung [15]. We assume a monotonic relationship between chest volume and lung volume [14], and by using a first order approximation, we have Vlung (t) = k1 Vchest (t) + C

(3)

where, k1 and C are the slope and offset of this linear relationship, and assumed to be constant for a given individual. Airflow, Q(t), can be approximated as the volume change in the chest or more precisely, the thoracoabdominal region per time unit. This is the derivative of chest volume, Vchest (t) over time. Using numerical techniques, the airflow is calculated as Q(t) =

k1 ΔVchest (t) dVlung (t) ≈ . dt Δt

(4)

Based on (4), we have Q(t) > 0 during inspiration and Q(t) < 0 during expiration. This results in an always-positive airway resistance, as defined in (1).

Fig. 2. Calibrated V chest (t) and estimated breathing rate using depth data from Kinect.

C. Breathing Rate Measurement The chest-wall movements correspond to the inspiration and expiration is clearly apparent in Vchest derived from (2). The peaks and the valleys of Vchest match the end of each inspiration and expiration, respectively. Thus, a conventional peak detection algorithm can indicate the start and end of each breathing cycle. The patient’s breathing rate can then be calculated in real time as the number of breaths in a 1 min moving window. By calculating the maximum Vchest changes over a single spontaneous breath, the tidal volume, which is the volume of air that is inhaled or exhaled in a single breath can be calculated. Tidal volume is an important parameter in respiratory medicine due to its role in mechanical ventilation [16]. For this research, we calibrated the volume calculation method for the VB-PARE volume output to be in liter. The calibration process consists of using a known-volume box, and compute its volume using (2) to find the scaling factor that converts the size of the box to the metric values based on the distance from the camera. Fig. 2 shows the calibrated Vchest (t) over 1 min of normal breathing in the sitting position. The breathing rate using depth data in this case was 18 breaths/min. IV. BREATHING EFFORT ESTIMATION USING PPG In this section, we show that ΔP can be estimated using PPG data. Studies have shown that breathing has a direct effect on pulse rate and blood pressure [13]. During inhalation, the expansion of the rib cage causes the pressure in the lungs and alveoli to become slightly negative relative to the atmospheric (min) pressure, as indicated by Plung . Similarly, the systolic blood pressure declines and pulse rate goes up slightly. On the other hand, when the volume of the thoracic cavity decreases during expiration due to respiratory muscle relaxation, the pressure inside the lungs becomes positive compared to the atmospheric (max) pressure, as indicated by Plung . Similarly the systolic blood pressure increases slightly. Pulsus Paradoxus (PP) is a parameter that indicates the changes in cardiac output caused by the changes in intrathoracic pressure, which also correlates with alveolar pressure, during breathing and corresponds to the force required for breathing [17]: (max) (min) P P ∝ ΔPlung = k2 Plung − Plung . (5)


1907

Fig. 3. High peaks and low peaks detected by Algorithm 1 in 1-min PPG waveform, which was recorded during normal breathing from a healthy participant.

If we assume inhalation and exhalation are symmetric, then (max)

(min)

Plung − Pair ≈ Pair − Plung , (max)

(min)

⇒ 2Pair = Plung + Plung . By combining (5) and (7): (min) P P = 2k2 Pair − Plung .

(6) (7)

(8)

Substituting (4) and (8) in (1) and combining k1 and k2 in k, (min) at the end of each inhalation, when Plung = Plung , we can link ΔVchest and P P to R R=k

PP Δt. ΔVchest (t)

(9)

A. Breathing Rate Measurement and PP Estimation PPG is a noninvasive measurement of blood flow with optical sensors placed at the surface of the skin (e.g., finger). Photoplethysmographic waveforms captured by pulse oximetry devices shows phasic variations related to the respiratory cycle [18]. Frey and Butt in [19] showed that PP is strongly correlated (with the correlation coefficient of 0.85) to the fluctuations in the PPG waveform during a breathing cycle. The PPG fluctuation is defined as the difference between the highest and lowest values of the upper peaks in the PPG waveform in a given breathing cycle. Fig. 3 illustrates the highest and lowest peaks of a 1-min normalized PPG recoding from a healthy individual in resting condition. To extract the PPG fluctuations, the recorded raw PPG was low-pass filtered with a fourth-order Butterworth filter with 4−Hz cutoff frequency, which was chosen based on the fact that normal breathing rate in adult is 12–20 breaths/mine ( 1 year > 1 year Same Same Same < 1 year NM Same > 1 year

2.38 1.51 1.62 2.9 2.18 2.05 1.92 2.41 1.06 1.76 2.48 2.26 2.36 2.34

4.01 3.48 1.85 3.81 2.23 3.1 1.99 2.85 1.75 2.2 3.6 3.19 3.68 2.75

NM 0.66 0.99 1.54 2.93 1.47 2.84 NM 0.98 2.26 2.01 NM 0.83 2.73

NM 4.14 3.03 5.36 5.16 4.46 4.41 NM 2.73 4.46 5.21 NM 4.51 5.47

NM 0.23 0.48 0.17 0.07 0.19 0.02 NM 0.28 0.18 0.28 NM 0.5 0.17

NM stands for not-measured plethysmography results.

the IRB inclusion/exclusion criteria, which included being diagnosed with a degree of airways obstruction, being able to sit still on a chair for at least 8 min, and being able to follow simple instructions, such as placing the index finger on the top of a SpO2 sensor. Participants’ demographic for Phase II, sorted based on gender and age, is presented in I. Participants were asked to wear fairly tight t-shirts during respiration monitoring in order to be consistent and minimize shape artifacts, such as wrinkles and loose parts on their clothing. B. Data Acquisition System A Kinect for Windows v2 sensor (Microsoft, Redmond, WA, USA) [20] and a PPG BME-Kit (PhysioLab, Korea) [21] were used for data acquisition in this study. The PPG BME-Kit is composed of a pulse oximeter sensor (one infrared channel) fixed on a printed circuit board (PCB) and its driving circuitry, which raw analog output was sampled at 1 kHz. Because the PPG BME-Kit sensor was fixed on PCB, the collected data was prone to motion artifacts. Therefore, in Phase II, a custom-built casing was used to reduce the finger movements on the sensor. Kinect sensor is comprised of an RGB camera, an infrared depth sensor, three-axis accelerometer, and a four-array microphone. For our experiment, using SDK 2.0, the depth sensing capability with the resolution of 512 × 424 pixels at a frame rate of 30 f/s was leveraged to measure chest-wall movements. The Kinect was mounted on a camera stand over (Phase I) or in front of (Phase II) the participants chest, such that the center of the depth sensor was aligned with the center of the participants chest with 130-cm distance between them, (as shown in Fig. 4. This distance was chosen based on the study results of [22], where the Alnowami et al. showed that for Kinect, an optimal performance with 1-mm depth resolution can be achieved at a distance between 85 and 120 cm from the camera. The 130 cm is chosen to take into consideration the width of the chest and the abdominal regions of the body.

Fig. 4.

Data collection setup in: (a) Phase I and (b) phase II.

five plastic straws bundled together in disposable mouthpieces [2]. Each straw was 65 mm in length and 3 mm in diameter. A picture of our custom-made mouthpieces with different numbers of straws can be seen in Fig. 4(a). Participants were asked to lie down in supine position on a yoga mat, facing the Kinect camera, and were asked to keep their index finger of the right hand on the PPG kit on the floor. The supine position was chosen to ensure no-motion on the participant’s back during breathing. Using one, three, and five plastic straws, participants were asked to simulate the obstructed spontaneous breathing, each for three trials of 1-min length, while wearing a nose clip. In Phase II, since lying down on the ground was not comfortable for some of the asthmatic patients, the participants were asked to sit in a comfortable posture on a chair in front of the Kinect and insert the right index-finger into our customdesigned PPG sensor. The participants were asked to breathe normally, while remaining as still as they could for ∼8 min. Fig. 4(a) and (b) shows the data acquisition setup in Phases I and II, respectively. VI. EXPERIMENTAL RESULTS

C. Data Collection Procedure

A. Phase I: Obstruction Prediction

In Phase I, different airway resistances were emulated by asking the healthy participants to breathe through one, three, and

Based on the Hagen–Poiseuille equation in fluid mechanics [23], relatively small changes in the radius of the airways can


1909

Fig. 5. Airway obstruction mapping based on the VB-PARE output OVB-PARE and the estimated airway resistance Raw .

cause significant changes in the airway resistance, even when simplifying the airways to cylindrical pipes. In the Phase I, starting with healthy participants who were experiencing normal airway resistance, we induced extra resistance externally, using a custom-made mouthpiece made of different numbers of straws [2]. Assuming Rs as the resistance induced by one straw, three parallel straws induce R3s , and five straws induce R5s . Note that these airway resistances were in series with the participants normal (i.e., low) airway resistance, and dominated overall airway resistance, while breathing through the mouthpiece. The objective in Phase I was to verify whether the VB-PARE system is capable of differentiating between different externally induced airway obstructions, defined in three levels as “Mild” from using five straws, “Moderate” from using three straws, and “Severe” from using only one straw. In other words, the airway obstruction was labeled according to Fig. 5, when the VB-PARE output, OVB−PARE , for N straws (N ∈ {1, 3, 5}) resulted in an airway resistance Raw in the following ranges: ⎧ 4 if Raw ≤ 15 Rs ⎪ ⎨ Mild, Airway 4 = Moderate, if 15 Rs < Raw ≤ 23 Rs (10) Obstruction ⎪ ⎩ Severe, if 23 Rs < Raw . VB-PARE system can generate several outputs including airflow, PP-correlated fluctuations, and also estimated resistance as given in (9). For any of these outputs, we assume a simplified linear relationship between the airway resistance Raw and the VBPARE output OVB−PARE , such as Raw = αOVB−PARE + β, where α and β are the slope and offset of this linear relationship, respectively. Having two unknowns α and β, we used two randomly chosen trials from one-straw and five-straw measurements to extract the linear relationship for each participant, and tested the accuracy of the prediction in the remaining trials for that given participant. The reported prediction accuracy in the next sections is averaged over all test sets, when different sets of training trials were chosen. For simplicity, Rs of one straw was normalized to one. 1) Obstruction Prediction Using Airflow: Reduced maximum airflow in spirometry is an indication of increased airway resistance and pulmonary obstruction [6]. Since our experiment

Fig. 6. (a) Airway obstruction prediction accuracy using airflow only (blue bars), PP values only (green bar), and a combination of both airflow and PP values (red bars). Asterisk (∗) indicate participants whose PPG reading had a high degree of artifacts. (b) RMSE between breathing rate estimated from Kinect versus PPG data.

only involves spontaneous breathing rather than forced breathing, we first used the extracted airflow, using the algorithm explained in Section III, as a measure for obstruction prediction. To preserve the linear relationship, we considered the inverse of the averaged maximum airflow during inspiration as the VB-PARE output OVB−PARE . In Fig. 6(a), the airway obstruction prediction accuracy for each of the healthy participant using only the airflow is shown in blue bars, resulting in an average predication accuracy of 70.9 ± 19.2%. The high standard deviation suggests that for some participants, the airflow was not sufficient for the prediction of the severity of the airway obstruction. This stems from the fact that by applying different effort levels during breathing, humans are able to temporarily adjust their respiration airflow, even when they have high airway obstruction. For instance, although in most healthy participants, the breathing rate dropped with increased induced resistance, participant #6, was able to keep the breathing rate approximately constant in the 14 breaths/min range for all of the trials even with one straw, while exerting more effort. 2) Obstruction Prediction Using Airflow and PP: In this section, we evaluate the effect of adding PP-correlated values extracted from PPG waveform, as explained in Section IV, as a measure of breathing effort to the airway obstruction prediction algorithm. PP-correlated value, used in the predication, was the

1910


Fig. 8. Lung volumes and capacities: The volume of air associated with different phases of the respiratory cycle.

Fig. 7. Samples of PPG readings, clean (top), corrupted by finger misplacement (middle), and corrupted by motion artifacts (bottom).

median of the “PP-correlated fluctuations” from all breaths in each trial for a given participant, computed using Algorithm 1. First, we calculated the prediction accuracy using only P P as the OVB−PARE , and then, we updated the predication by combining P P and airflow as directed in (9), and used the estimated resistance as the OVB−PARE . The green bars in Fig. 6(a) indicate the airway obstruction prediction accuracy based on P P only, resulting in average predication accuracy of 45.9 ± 13.9%. This result proved that breathing effort measured by P P is not a reliable measure of airway obstruction by itself. However, using both P P and Q, the average accuracy of the obstruction predication, indicated by the red bars in Fig. 6(a) was increased to 76.2 ± 13.8%. While, the predication accuracy increased in most of the participants, for participants #4, #8, #9, and #11, shown by ∗ in the Fig. 6(a), this was not the case. Looking more carefully at the PPG readings from these participants, we found that in some cases the PPG recordings were corrupted by artifacts due to finger movements during the trial, as well as misplaced finger on the PPG board, as shown in Fig. 7 examples. As stated in [24], the performance of a lot of portable and wearable biosensors including PPG sensors is highly influenced by motion artifacts. Intelligent postprocessing algorithms can be applied to reduce the signal corruption due to these artifacts [25]. The participants’ breathing rate can be extracted from either depth or PPG data from Kinect or SpO2 sensors, as explained in Sections III-C and IV-A, respectively. Given the high possibility of motion artifact in PPG recording, we used the comparison between breathing rates to detect cases of corrupted PPG recordings due to artifacts. Fig. 6(b) shows the root-mean-square error (RMSE) between the breathing rates obtained from depth and PPG data for each Phase-I participant, resulting in an average RMSE of 3.1 ± 2.6 breath/min among all the participants. Based on this observation, for Phase II, we changed the data collection

Fig. 9. (a) Breathing rates extracted from depth data (blue bars) and PPG data (red bars). (b) Tidal volumes extracted from VB-PARE (blue bars) and plethysmography (red bars). NM stands for participants with no plethysmography measurement results.

procedure by adding a custom-built casing to the PPG board that held the participants finger steady on the pulse oximeter sensor. B. Phase II: Comparison With PFT Results All participants in the Phase II had already performed spirometry test at Emory Clinic. Moreover, 11 out of 14 participants had completed the whole-body plethysmography test to assess the severity of their airway obstruction by measuring their lung volumes and capacities, as defined in Fig. 8. I shows their summary of PFT results along with the time interval between spirometry and plethysmography tests for each patient participant. The longer these two tests are apart from each other, the higher the possibility of the outcome mismatch due to the changes in patient’s pulmonary function over time, including the response to treatments. To validate the quality of the PPG data in Phase II, the estimated breathing rates from Kinect and SpO2 were compared, resulting in Fig. 9(a), which indicates much smaller average mismatch (1.4 ± 1.5 breath/min) between breathing rates extracted from these two measures, due to less corrupted PPG data. The highest mismatches of ∼5 breaths/min were from participants #12 and #14, who were the oldest. Visual inspection of the PPG


1911

TABLE II CLASSIFICATION OF THE AIRWAY OBSTRUCTION SEVERITY

Severity

FEV1/FVC (%)

FEV1 (% predicted)

sGaw (L/s/cmH2 O/L)

Mild Moderate Severe

< 70 < 70 < 70

≥ 80 [50–80) [30–50)

− [0.2–0.15) ≤0.15

data revealed high-motion artifacts, which might be the result of the participants’ hand tremor. 1) Tidal-Volume Comparison: Healthy adults have an average tidal volume of 0.5 liter. Both VB-PARE and plethysmography can provide the average tidal volume during spontaneous breathing. To evaluate the accuracy of the respiration monitoring algorithm in Section III, Fig. 9(b) compares the tidal volumes calculated using VB-PARE and plethysmography in blue and red bars, respectively. Correlation coefficient between the two measurement methods was 0.77 for 95% confidence interval with an average error of 0.07 ± 0.06 liter. The majority of the tidal volumes from VB-PARE are smaller than plethysmography, which might be from the fact that Kinect records only the chest movements during breathing, whereas lung plethysmography measures volume changes in all directions. In future research, this small discrepancy can be corrected by mapping the volumes measured by Kinect (Vchest ) to the volumes measured by lung plethysmography (Vlung ) using (3). The approximation parameters in (3), can be estimated using an accurate flowmeter, while breathing in front of the Kinect camera. The highest mismatch (0.2 liter) was recorded from participant #3 with the BMI of more than 50. This participant had a shallow but a fast (28.1 breaths/min) breathing pattern, which is considered inefficient breathing, resulting in a condition, known as obesity hypoventilation syndrome, in which poor breathing leads to lower oxygen and higher carbon dioxide levels in the blood [26]. 2) Spirometry Versus Plethysmography: Physicians combine the PFT results from spirometry, including the forced expiratory volume in 1 s (FEV1), and the ratio between the FEV1 and forced vital capacity (FVC) as two objective measures to assess the airway obstruction severity. Table II shows different stages based on the chronic obstructive pulmonary disease severity classification, adopted from Global Initiative for Chronic Obstructive Lung Disease [27], where predicted FEV1 is the normal value in a healthy individual with the same height, weight, and race. On the other hand, the whole-body plethysmography calculates specific airway conductance (sGaw) during tidal breathing and the sGaw values less than 0.15–0.2 L/sec/cm H2 O/L are consistent with airway obstruction [28]. Using these numbers, we defined the severity stages as listed in Table II. Out of the 14 participants in phase II, five had FEV1/FVC