Upper airway effective compliance during wakefulness and sleep in obese adolescents studied via two-dimensional dynamic MRI and semiautomated image segmentation

Abstract

Novel biomarkers of upper airway biomechanics may improve diagnosis of obstructive sleep apnea syndrome (OSAS). Upper airway effective compliance (EC), the slope of cross-sectional area versus pressure estimated using computational fluid dynamics (CFD), correlates with apnea-hypopnea index (AHI) and critical closing pressure (P_crit). The study objectives are to develop a fast, simplified method for estimating EC using dynamic MRI and physiological measurements and to explore the hypothesis that OSAS severity correlates with mechanical compliance during wakefulness and sleep. Five obese children with OSAS and five control subjects with obesity aged 12–17 yr underwent anterior rhinomanometry, polysomnography, and dynamic MRI with synchronized airflow measurement during wakefulness and sleep. Airway cross section in retropalatal and retroglossal section images was segmented using a novel semiautomated method that uses optimized singular value decomposition (SVD) image filtering and k-means clustering combined with morphological operations. Pressure was estimated using rhinomanometry Rohrer’s coefficients and flow rate, and EC was calculated from the area-pressure slope during five normal breaths. Correlations between apnea-hypopnea index (AHI), EC, and cross-sectional area (CSA) change were calculated using Spearman’s rank correlation. The semiautomated method efficiently segmented the airway with average Dice Coefficient above 89% compared with expert manual segmentation. AHI correlated positively with EC at the retroglossal site during sleep (r_s = 0.74, P = 0.014) and with change of EC from wake to sleep at the retroglossal site (r_s = 0.77, P = 0.01). CSA change alone did not correlate significantly with AHI. EC, a mechanical biomarker which includes both CSA change and pressure variation, is a potential diagnostic biomarker for studying and managing OSAS.

NEW & NOTEWORTHY This study investigated the dynamics of the upper airway at retropalatal and retroglossal sites during wakefulness and sleep by evaluating the effective compliance (EC) of each site and its correlation with apnea-hypopnea index (AHI) using novel semiautomated image processing. AHI correlated significantly with retroglossal EC during sleep and change of retroglossal EC from wake to sleep. The results suggest EC as a promising noninvasive diagnostic marker for estimating the mechanical properties of various upper airway regions in patients with OSAS.

INTRODUCTION

Obstructive sleep apnea syndrome (OSAS) affects 2%–3% of children worldwide. Children with obesity face 50% higher prevalence in developing OSAS (1). Treatment with adenotonsillectomy (AT) for OSAS has proven less effective in children with obesity compared with normal weight children (2–4). Hence, to understand the mechanism underlying the high prevalence of OSAS and lower effectiveness of AT for children with obesity, many recent studies have provided insight into functional mechanisms that could promote OSAS and contribute to poor AT response in obese children with OSAS. Previously, we demonstrated that analysis based on magnetic resonance (MR) images can lead to better understanding of airway biomechanical properties of children with OSAS (5–7). Persak et al. (5) estimated the pressure field in the upper airway during sedative-induced sleep and found that the effective airway compliance (EC), which is the slope of airway cross-sectional area versus airway pressure curve, was higher in young children with OSAS than in control subjects. Image-based computational fluid dynamics (CFD) techniques and nasal resistance were used to study pharyngeal mechanical compliance of obese adolescent girls with OSAS (1). EC in the nasopharynx was often negative in awake children with OSAS, and lower than in control subjects. EC was also found to be inversely correlated with apnea-hypopnea index (AHI) and critical airway closing pressure, P_crit. In summary, EC is a potential diagnostic marker based on localized biomechanics that may be related to apnea severity in children.

There are several challenges to estimating EC using the current image-based CFD method that may limit its wider application in clinical research or diagnostics. Image-based airway CFD is time-consuming, usually requiring several days of technician time per subject. It requires specialized and often expensive software and high-performance computers, engineering expertise, and advanced imaging methods such as retrospective gating during tidal breathing. The primary aim of this work is to demonstrate a simpler non-CFD method to estimate EC based on continuous dynamic two dimensional (2-D) MRI that may allow a technician to quickly quantify airway mechanics, for both tidal and abnormal or obstructed breaths.

To demonstrate the new method, the cross sections of the airway at retropalatal and retroglossal sites were segmented from 2-D dynamic MRI to investigate the EC of the airway in natural sleep and awake states, based on nasal resistance, and flow rate and nasal pressure data synchronized to the MR images. Bitners et al. (8) studied the effect of sleep on upper airway dynamics and found that the percent change of cross-sectional area (CSA) in OSAS subjects with obesity during sleep was significantly higher than in control subjects with obesity, and significantly higher in sleep than during wakefulness. The second objective of this study is to investigate the relationship between upper airway EC at retropalatal and retroglossal sites during wakefulness and sleep and OSAS severity as measured by the apnea-hypopnea index (AHI). We hypothesize that EC, which includes variation in both area and pressure load during one or more respiratory cycles, is a biomechanical indicator of upper airway pressure sensitivity that will correlate with AHI.

METHODS

Subjects

This clinical study was conducted by Children’s Hospital at Montefiore and approved by the Institutional Review Board at the Albert Einstein College of Medicine. Subjects in this study were children 12–17 yr of age, who underwent upper airway dynamic MRI during both wakefulness and natural sleep and independent overnight polysomnography to detect and quantify OSAS. Informed consent was collected from the participants or parents of minors at enrollment.

The present study includes 10 subjects (4 females and 6 males). Half of the subjects were diagnosed with OSAS (3 females and 2 males), and control subjects were selected to match the OSAS subject demographic characteristics. The subjects were included based on the following criteria:

1. Awake and spontaneous natural sleep achieved in the MRI scanner during MRI protocol.

2. Nasal breathing during the scans.

3. Sufficient MR image quality with visible airway and without artifact to segment airway.

4. Rhinomanometry data without obstruction of either nasal passage, and three or more consistent breaths.

Participants are part of a larger study investigating upper airway biomechanical characteristics in adolescent children with and without OSAS. Subjects and data in this study have been used in the prior publication of airway cross-sectional variation in OSAS of Bitners et al. (8), which used different methods of analysis that do not require rhinomanometry data. The analysis presented here is on a smaller subset of participants that met the more restrictive criteria mentioned earlier and were suitable for EC calculation.

Polysomnography

Overnight polysomnography (Natus/Xltek, Oakville, ON, Canada) was performed at the Sleep Disorders Center at the Children’s Hospital at Montefiore. Sleep was staged and respiratory events were scored based on standard criteria (9). The definition of hypopnea adopted was 30% decrease in amplitude of the oronasal thermal sensor (compared with tidal breathing) associated with either a decrease in basal $\mathrm{S}\mathrm{p}\mathrm{O}{}_2$ by ≥3% or an arousal, and an obstructive apnea was defined by at least a 90% decrease in peak amplitude of the oronasal thermal sensor associated with continued respiratory effort lasting longer than two respiratory cycles. OSAS was diagnosed if subjects had obstructive apnea index > 1 event/h or AHI > 5 events/h, or both.

MRI Protocol

The MRI protocol was described previously in Bitners et al. (8). In total, 300 dynamic images, or frames, were taken at retropalatal and retroglossal sites over a duration of 1.6–1.8 min during both wakefulness and sleep. The retropalatal imaging plane bisected the soft palate perpendicular to the airway; the retroglossal imaging plane was located between the uvula and epiglottis and perpendicular to the airway. Sleep was determined by behavioral and physiologic criteria including absent response to a voice command (patient name call) and physiological changes (presence of snoring, irregular flow signal/flow limitation, or drop in mask pressure with or without 3% oxygen desaturation). In addition, the presence of sleep was verified at the end of the study by confirmation with the patient (8). Nasal-oral airflow volume, pressure, oxygen saturation, and MRI gating pulses were recorded synchronously with MRI data acquisition (8).

Rhinomanometry

Anterior rhinomanometry (NR6 Rhinomanometer, GM Instruments) was conducted on the subjects upon enrollment at Children’s Hospital at Montefiore to evaluate nasal resistance and nasal flow, as previously described in Sin et al. (6). Rohrer’s coefficients were extracted from inspiratory and expiratory phases for left and right nasal passages (7). Breaths with peak pressure above 150 Pa were used for all measurements.

Semiautomated Segmentation of the Cross Section of the Dynamic Airway

The two main variables needed to investigate effective compliance of a subject’s upper airway at a cross-sectional plane are the airway area and airway pressure at that plane. To efficiently estimate the cross-sectional area (CSA) of the airway from the MR images taken at a plane in retropalatal and retroglossal sites, a semiautomatic segmentation method was developed (23). In this method, the images are first filtered and then segmented for the CSA semiautomatically.

The dynamic MR images were filtered using singular value decomposition (SVD) implemented in MATLAB (The MathWorks, Inc.) (10). For each acquired study for the two sites, the m dynamic images, each with n pixels, are stored as the row vectors of a matrix I of size m × n. The filtered images, I_r, are obtained from the truncated SVD of I:

$$I_r=U_r{S}_r{V}_r^T,$$ (1)

where U_r and V_r are orthonormal matrices of size m × r and n × r, respectively, S_r is a diagonal matrix containing the first r singular values of I, and r is the rank used for the filter. The choice for the rank r is often based on visual inspection of the filtered image; in general, a small rank yields smoothed images, and a higher rank includes more granularity. We tested two methods by Epps et al. (10–14) to automatically determine the rank for the filter. A typical case is illustrated in Fig. 1. Figure 1A shows an unfiltered MR image, Fig. 1B) shows its low-rank filtered image (r = 2) using the method detailed in Epps et al. (14), and Fig. 1C shows the filtered image (r = 38) obtained by the “E15” method (10–13). The “E15” method for setting the rank provides a more accurate reconstruction of the original image than the low-rank filter and is used for filtering of the MR images.

Figure 1.

Singular value decomposition (SVD) filtering. An original retropalatal MR image at retropalatal site (A); its corresponding low-rank filtered retropalatal MR image (r = 2) (B); and its rank-optimized “E15” filtered retropalatal MR image (r = 38) (C). MR, magnetic resonance.

A bounding box was fixed manually to contain the airway throughout the 300-frame images. This was the only manual step (other than quality checks) in the segmentation process. The airway cross section was segmented in the filtered image within the bounding box for each time frame using an adaptive k-means clustering algorithm (10, 15–17). The first cluster, with the lowest intensity, was interpreted as the airway.

Subsequently, nonairway pixels were removed and small holes within the segmented region were filled by using morphological erosion and dilation operations (10, 17) available in MATLAB (The MathWorks, Inc.) (10). The algorithm consisted of the following steps:

1. Remove small four connected components of pixels (i.e., pools of pixels that are connected strictly by edges) having <30 pixels in the component.

2. Erosion followed by dilation using a 2 × 2 pixel structuring element, to remove very narrow components and disconnect components connected by long, narrow connections.

3. Repeat step 1.

4. Calculate the cross-sectional area of the airway.

This process is illustrated in Figure 2.

Figure 2.

A: “E15” filtered retropalatal MR image with a manually set bounding box (right hand side—magnified) containing the airway. B: the pixels inside the bounding box labeled with different cluster indices after performing k-means clustering with k = 3. C: the final segmented airway (white), after morphological and connected component operations. MR, magnetic resonance.

The segmentation result in each frame of each sequence is checked visually by overlaying a transparent-colored mask of the airway cross section on the original airway image. Problems identified during visual checking were solved by adjusting bounding box location or the cluster number.

Comparison of automated segmentation with ground truth segmentation.

The automatic segmentation method was evaluated by comparison of the output with ground truth segmentation for both airway sites and sleep/wake states. The ground truth segmentation was performed manually at each protocol (retropalatal-wake, retropalatal-sleep, retroglossal-wake, and retroglossal-sleep) by a trained biomedical engineer. The commonly used Dice Coefficient (DC) (18), which expresses the fractional spatial overlap between the semiautomated segmentation and ground truth, was used for evaluating the quality of segmentation:

$$DC(A_g,A_p)=\frac{2\left(\right|A_g\cap A_p\left|\right)}{\left|A_g\cup A_p\right|}=\frac{2\left|TP\right|}{2\left|TP\right|+\left|FP\right|+\left|FN\right|},$$ (2)

where A_g and A_p are the sets of ground truth and segmented pixels, respectively, and |x| denotes the cardinality of set x. TP, FP, and FN are sets of true positive, false positive, and false negative pixels, respectively (Fig. 3).

Figure 3.

Example illustrating segmentation evaluation. A: ground truth segmentation overlaid in red on a retropalatal MRI frame. B: segmented cross section overlaid in green. C: the airway masks superimposed, showing true positive (black), false negative (red), and false positive (green) regions; DC = 93.6% for this frame. DC, Dice Coefficient.

Airway Pressure

Although the shape of the pharynx causes significant pressure variation in the airway, it is quite common that this pressure variation is relatively small compared with the pressure drop through the nasal passages. A review of CFD models of tidal breathing in awake adolescent girls from Wootton et al. (1) showed that pressure fluctuation at the choanae was within 20% of pressure fluctuation at the narrower retropalatal pharynx in 75% of the subjects, and within 5% in 50% of the subjects. Therefore, the pressure at the choanae is often a good indicator of the pressure in the pharynx and, as it can be quickly estimated using flow rate and nasal resistance, it was used in this method.

To account for the nonlinear relationship between flow and pressure drop through the nasal passages, choanae pressure was calculated using the Rohrer’s equation for each nasal passage (19). Time-varying airway pressure was estimated using frame-averaged inflow rate, and nasal resistance Rhorer’s coefficients of left and right nasal passages. Airway pressure and volumetric flow rate of left and right nasal passage were calculated by solving the following simultaneous equations:

$${\dot{V}}_T={\dot{V}}_L+{\dot{V}}_R,$$ (3)

$${\mathrm{P}}_{\mathrm{a}\mathrm{i}\mathrm{r}\mathrm{w}\mathrm{a}\mathrm{y}}\approx {\mathrm{P}}_{\mathrm{c}\mathrm{h}\mathrm{o}\mathrm{a}\mathrm{n}\mathrm{a}\mathrm{e}}={\mathrm{P}}_{\mathrm{m}\mathrm{a}\mathrm{s}\mathrm{k}}-K_1{,}_L{\dot{V}}_L-K_2{,}_L{\dot{V}}_L\left|{\dot{|V}}_L|\right|,$$ (4)

$${\mathrm{P}}_{\mathrm{a}\mathrm{i}\mathrm{r}\mathrm{w}\mathrm{a}\mathrm{y}}\approx {\mathrm{P}}_{\mathrm{c}\mathrm{h}\mathrm{o}\mathrm{a}\mathrm{n}\mathrm{a}\mathrm{e}}={\mathrm{P}}_{\mathrm{m}\mathrm{a}\mathrm{s}\mathrm{k}}-K_1{,}_R{\dot{V}}_R-k_2{,}_R{\dot{V}}_R\left|{\dot{|V}}_R|\right|,$$ (5)

where ${\dot{V}}_T$, ${\dot{V}}_L$, and ${\dot{V}}_R$ are the total frame-averaged measured flow rate and flow rates through left and right nasal passages, respectively. P_airway, P_choannae, and P_mask are the airway pressure, choanae pressure, and pressure measured in the subject’s mask, respectively, and K_1,L, K_2,L, K_1,R, and K_2,R are the measured nasal resistance Rohrer’s coefficients of left and right nasal passages.

The error due to using the choanae pressure (Eqs. 4 and 5) instead of local pressure was estimated using CFD in one control subject and all five OSAS subjects (IDs 55, 64, 65, 86, 90, and 102). One frame from a three-dimensional (3-D) dynamic MR image sequence was selected from a tidal breath at peak inspiratory flow, segmented, and a CFD model constructed using previously described methods (1) to calculate the pressure distribution in the pharynx during peak inspiratory flow. The average pressure at each retropalatal and retroglossal imaging plane was computed, and the relative errors of using the choanae pressure to estimate the retroglossal pressure averaged 10% (1%, 6%, 8%, 25%, 18%, and 3%, respectively) for these six subjects. The differences for the retropalatal pressure were similar but up to 2% smaller in each case.

Dynamic Cross-Sectional Area Change and Effective Compliance Estimation

For each subject, the CSA and pressure relationships at retropalatal and retroglossal sites of the airway during wakefulness and sleep were studied for five continuous artifact-free tidal breaths with stable pressure and without swallowing (8) (Fig. 4A). The dynamic CSA change over the five breaths was defined as follows:

$$\%CSA=\frac{CS{A}_{\max \hspace{0.17em}}-CS{A}_{\min \hspace{0.17em}}}{CS{A}_{\min \hspace{0.17em}}}\times 100\%,$$ (6)

where CSA_max and CSA_min are the maximum and minimum area over the selected five breaths.

Figure 4.

A: area and pressure vs. time (first 5 consecutive tidal breaths shown in box). B: cross-sectional area vs. pressure in an awake subject with OSAS (AHI = 9.3); EC = 0.2014 mm²/Pa. AHI, apnea-hypopnea index; EC, effective compliance; OSAS, obstructive sleep apnea syndrome.

EC was calculated from the average slope of the pressure versus dynamic area plot estimated for five breaths (typically consisting of 30– 45 data points), using least squares error minimization (Fig. 4B). Positive slope (positive EC) indicates cross-sectional area decreased during inspiration, indicating net passive behavior, whereas negative EC indicates area expanding during inspiration, indicating net active behavior.

Statistical Analysis

Statistical analyses were conducted using Microsoft Excel and MATLAB. Sample mean and standard deviation were tabulated for all variables. OSAS and control group demographics were compared using two-tailed t tests. Spearman’s correlation coefficient (r_s) was computed between effective compliance and AHI using pooled control and OSAS subject data sets.

Uncertainty of the cross-sectional area was estimated as the root-mean-square error of the automated segmentation relative to the ground truth segmentation. Uncertainty of the effective compliance was estimated by mean square error propagation of flow rate (5%) and nasal resistance (2%) measurement instrument uncertainties, and the uncertainty of the change in area.

RESULTS

Subject Demographics

Subject demographics are summarized in Table 1. All subjects were obese; the body mass index of the subjects ranged from 29.26 to 41.40 kg/m² and BMI Z scores were above 1.9. There was no statistically significant difference between the two groups in age, sex, weight, BMI, or height. One subject in the OSAS group had mild to moderate AHI (9.1), the other four subjects had much more severe OSAS, with AHI ranging from 17.8 to 37.2 events/h.

Table 1.

Demographics

	Controls	OSAS	P Value
n	5	5
Age, yr	14.3 ± 1.5	15.3 ± 1.9	0.37
Sex, male/female	4/1	2/3	0.24
Height, cm	164.6 ± 7.8	167.0 ± 15.7	0.77
Weight, kg	89.3 ± 13.1	99.3 ± 26.1	0.487
Body mass index, kg/m2	32.8 ± 2.8	34.3 ± 4.1	0.52
Apnea hypopnea index	2.1 ± 1.5	25.4 ± 11.8	0.011

Values are means ± SD; n, number of subjects.

Semiautomated Segmentation Method

The Dice Coefficient (DC) shows the benefits of the different image processing steps to improve semiautomated segmentation accuracy relative to ground truth manual segmentation. Optimal filtering increased DC in 94% of the images, and increased average DC from 86.85% to 90.39% (Fig. 5). Filtering also greatly increased segmentation accuracy for difficult images; the minimum DC increased from 44% to 71%. For the k-means clustering image processing step, increasing the number of clusters k from two to three improved DC as the average DC increased from 85.84% to 89.41%. But further increases in k had minimal effect on DC (<1% improvement) as shown in Fig. 6, whereas the computational cost increased significantly. Therefore, k = 3 was chosen as the optimal value.

Figure 5.

Dice Coefficient of sample MR image set before (original) and after optimal E15 filtering (filtered). MR, magnetic resonance.

Figure 6.

Effect of varying the number of clusters k in k-means clustering on Dice Coefficient.

DC ranged from 82% to 98%; average DC for five continuous artifact-free tidal breaths for the four protocols, retropalatal-wake, retropalatal-sleep, retroglossal-wake, and retroglossal-sleep, was 92.63%, 89.21%, 93.58%, and 89.95%, respectively (Fig. 7). The uncertainty of the area calculated by the semiautomated segmentation method was 13%.

Figure 7.

Frame by frame DC: retropalatal airway (A) and retroglossal airway (B) from one control subject (AHI = 0.7). AHI, apnea-hypopnea index. DC, Dice Coefficient.

Dynamic Cross-Sectional Area Change and Effective Compliance

The dynamic cross-sectional area change (%CSA) results are shown in Fig. 8. At the retropalatal site, %CSA increased during sleep by 40% in three subjects, and was similar or decreased in the remaining seven (Fig. 8A). At the retroglossal site, %CSA increased during sleep by 40% or more in seven subjects (Fig. 8B). The largest %CSA increase at both sites is in an OSAS subject. Compared with Bitners et al. (8), this result is consistent at the retroglossal site but not confirmed at the retropalatal site, perhaps due to the smaller sample size in the present study.

Figure 8.

Cross-sectional area variation during tidal breathing in all subjects. A: %CSA at retropalatal site of upper airway during wakefulness and sleep. B: %CSA at retroglossal site. %CSA, cross-sectional area percent change; OSAS, obstructive sleep apnea syndrome.

The ECs of control and OSAS subjects are presented in Fig. 9. At the retropalatal site, four subjects show larger EC change > 0.05 mm²/Pa from wake to sleep (Fig. 9A). EC decreased with sleep in three of these four subjects. Furthermore, two OSAS and one control subject had negative EC both during wakefulness and sleep. At the retroglossal site, 6 out of 10 subjects show EC change > 0.05 mm²/Pa from wake to sleep, with five subjects showing increased EC (Fig. 9B). Overall, more negative EC is observed at retropalatal site compared with retroglossal. This suggests a higher degree of phasic activation of pharyngeal dilators at the retropalatal site.

Figure 9.

Effective compliances of all subjects. A: EC at retropalatal site of upper airway during wakefulness and sleep. B: EC at retroglossal site. EC, effective compliance; OSAS, obstructive sleep apnea syndrome.

Correlation Analysis

To examine the relationship between AHI and biomechanical parameters [EC and %CSA during wakefulness and sleep, the change of EC (ΔEC) from wake to sleep, ΔEC and change of %CSA (Δ%CSA) from wake to sleep], Spearman’s rank correlation analysis was performed; the results are summarized in Table 2. AHI correlates significantly with retroglossal EC during sleep (r_s = 0.74 and P = 0.014), and with retroglossal ΔEC (r_s = 0.77, P = 0.010). Both relationships are plotted in Fig. 10. Retropalatal EC during wakefulness has a trend (not significant) toward negative correlation with AHI (r_s = −0.49, P = 0.15), which is consistent with previous CFD-based study on the nasopharynx (1). Furthermore, retroglossal ΔEC is strongly and positively correlated with retroglossal %CSA during sleep and retroglossal Δ%CSA (r_s = 0.77, P = 0.014 and r_s = 0.81, P = 0.0082, respectively). But neither retroglossal %CSA during sleep nor retroglossal Δ%CSA correlate significantly with AHI in this small sample, suggesting EC is more sensitive to OSAS severity than %CSA as a biomechanical marker.

Table 2.

Spearman’s rank correlations

	r s	P
AHI vs.
Retropalatal
EC-wake	−0.49	0.152
EC-sleep	0.22	0.53
ΔEC	0.32	0.37
Δ%CSA	0.26	0.48
Retroglossal
EC-wake	−0.22	0.53
EC-sleep	0.74	0.0141
ΔEC	0.77	0.010
Δ%CSA	0.43	0.22
Retroglossal ΔEC vs.
Retroglossal
%CSA-sleep	0.76	0.014
Δ%CSA	0.81	0.0082

Figure 10.

Scatter plots of AHI and effective compliance. A: AHI vs. retroglossal EC during sleep. B: AHI vs. retroglossal ΔEC. AHI, apnea-hypopnea index; EC, effective compliance.

DISCUSSION

In this study, dynamic MR images are used to quantify 2-D airway cross section and motion at retropalatal and retroglossal levels, which are commonly known sites of upper airway obstruction for children with OSAS (20). With rhinomanometry and synchronized flow rate data, the airway pressure is estimated, and this combination provides an opportunity to study upper airway biomechanics, in this case using EC. We found that AHI is strongly correlated to both retroglossal EC during sleep and ΔEC between wake and sleep.

Most of the retroglossal ECs in this study were positive both during wakefulness and sleep, but this was not the case at the retropalatal site. These results are consistent with our previous findings of negative EC in the nasopharynx (near the retropalatal site of this study) and more positive EC in the retroglossal oropharynx in a 3-D study of awake subjects (1). Positive EC indicates that the cross-sectional area becomes larger during expiration with positive pressure loads enlarging the airway (i.e., a passive airway) whereas negative EC indicates that the cross-sectional area becomes larger during inspiration in spite of negative pressure loads (i.e., an active airway). The EC results suggest different degrees of phasic activation of pharyngeal dilators in response to inspiration and expiration, and that the retropalatal site is more active than retroglossal. Both EC and %CSA are usually greater at the retroglossal site than retropalatal site, due to the typically larger retroglossal cross-sectional area. The smaller airway cross-sectional area at the retropalatal site adds challenge to MRI protocols and segmentation in some subjects, although the semiautomated segmentation algorithm performed well based on the DC (Fig. 5).

At the retroglossal site, the larger increase of EC and %CSA during natural sleep in OSAS subjects is consistent with studies examining the effect of sleep on upper airway dynamics (8) and airway dilator muscle electromyogram (EMG) (21, 22). The increase suggests lower tonic airway stiffness in OSAS subjects during sleep compared with control subjects, and relative phasic airway dilator inactivity during inhalation.

AHI was found to be positively correlated to retroglossal EC during sleep and change of EC from wake to sleep. This result is consistent with a model of a more compliant relaxed airway in OSAS subjects. In other words, the upper airway in OSAS subjects is compensated during wakefulness by higher phasic activation of airway dilator muscles.

The correlations between EC and AHI create the possibility to improve diagnosis for OSAS children. The attractiveness of this method is that it is noninvasive and can be performed during wakefulness and sleep. Even though retroglossal %CSA change is significantly and positively correlated to the change of EC from wake to sleep, the correlation of CSA with AHI did not reach statistical significance in this small group of subjects. This suggests that the biomechanical basis of EC, in this case using only nasal resistance and flow rate to estimate the internal pressure loads, may add greater sensitivity to measure the severity of OSAS compared with direct measurement of %CSA change alone.

There are several limitations to the present study. First, the sample size is small and may result in inconclusive correlations and group comparisons. The subject pool in Bitners et al. (8) is already limited because natural sleep during the MRI protocol is a challenge, and subjects who could not achieve sleep had to be excluded. Second, only five artifact-free continuous tidal breaths are selected and studied to estimate EC, which may lead to bias in breath selection as EC value may differ if a different set of breaths is used in the calculation. Subjects with high nasal resistance during rhinomanometry were excluded because high Rohrer’s coefficients give unrealistically high airway pressure extremes; high Rohrer’s coefficients usually occurred due to obstruction of one nasal passage, which interferes with the anterior rhinomanometry measurement. Third, the subject pool was limited to older children ages 12–17 yr. The method should also be applicable to adults and to younger children, and future studies should be conducted to validate the method in these populations. Finally, the semiautomated segmentation method relies on the manually set bounding box around the airway region on MRI and it is sensitive to the morphological operation for final airway CSA calculation. However, the semiautomated segmentation method is fast, and allows analysis of much longer sequences of images than are possible with manual segmentation, and future work will investigate hypopnea and apnea breaths, and tidal breathing trends before obstructive events.

Compared to previous studies using CFD to estimate the airway pressure and estimate the compliance (1, 5), this method is less accurate because the influence of the pharynx shape on the airway pressure is ignored. Ignoring the pharynx shape can create a large error in the pressure estimate in subjects with low nasal resistance or very restricted pharynx cross section. Considering the differences between local pressure CFD estimates and the choanae pressure used in this study, EC would be overestimated by 1%–25% (10% on average), and although the errors were not negligible, this method could rapidly give insight into the airway behavior. As more automated 3-D segmentation methods for airway MRI are developed, the accuracy of the pressure can be improved by incorporating the axial distribution of the pharynx cross-sectional area into a one-dimensional hydraulic model. The main advantages of the method are that it could be implemented by a technician without CFD training, and it is much less computationally intensive than CFD, and therefore potentially much more clinically feasible to give an indication of pharyngeal biomechanics.

CONCLUSIONS

In conclusion, we propose a time-efficient semiautomated method to study EC of the upper airway with 2-D dynamic MR images. The changes of dynamic CSA and EC at the retroglossal site are typically larger than at the retropalatal site. Retroglossal EC during sleep in OSAS subjects is higher than during wakefulness and the change of EC from wake to sleep is higher than in control subjects. AHI correlates significantly with retroglossal EC during sleep and the change of EC from wake to sleep. These promising results and the relative simplicity of the method suggest that EC is a promising diagnostic parameter for studying and managing the mechanical properties of various upper airway regions in patients with OSAS.

GRANTS

This work was funded by National Heart, Lung, and Blood Institute Grant R01 HL130468.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

D.M.L., M.E.W., R.A., and D.M.W. conceived and designed research; S.S. and M.E.W. performed experiments; K.R.C., R.A., S.S., and D.M.W. analyzed the data; K.R.C., S.S., Y.T., J.K.U., M.E.W., R.A., and D.M.W. interpreted results of experiments; K.R.C. prepared figures; K.R.C. drafted manuscript; K.R.C., S.S., Y.T., J.K.U., D.M.L., M.E.W., R.A., and D.M.W. edited and revised manuscript; K.R.C., S.S., Y.T., J.K.U., D.M.L., M.E.W., R.A., and D.M.W. approved the final version of the manuscript.