An Examination of Methodological Paradigms for Calculating Upper Airway Critical Pressures during Sleep

Abstract

Study Objectives:

The goal of this study was to examine different paradigms for determining critical closing pressures (Pcrit). Methods of determining Pcrit were compared, including direct observation of occluded (no flow) breaths versus inferring Pcrit from extrapolated data, and Pcrit generated by aggregating pressure-flow data from multiple runs versus Pcrit averaged across individual pressure-flow runs. The relationship between Pcrit and obstructive sleep apnea (OSA) was examined.

Methods:

A total of 351 participants with and without OSA underwent overnight polysomnography with pressure-flow measurements to determine Pcrit. A series of filters were applied to raw data to provide consistent, objective criteria for determining which data to include in Pcrit calculations. Observed Pcrit values were computed as the mean nasal pressure level at which a subject had at least two breaths with peak inspiratory flow < 50 mL/sec. Extrapolated Pcrit was calculated in two ways: (1) separately for each individual run and then averaged; and (2) using all valid data from individual runs combined into one plot.

Results:

Observed Pcrit was calculated in 67% to 69% of participants, a similar or higher proportion of study subjects compared to extrapolated Pcrit values using a ± 3 cm H2O filter. Although raw (unfiltered) extrapolated Pcrit measures were able to be calculated among a greater proportion of participants than filtered, extrapolated Pcrit values, and thus had fewer missing values, they had larger variability. Both extrapolated and observed Pcrit were higher among individuals with OSA compared to those without OSA.

Conclusions:

Observed Pcrit provides a reliable descriptor of hypotonic upper airway collapsibility. Different methods for determining Pcrit were able to distinguish subjects with and without OSA.

Citation:

Pien GW, Keenan BT, Marcus CL, Staley B, Ratcliffe SJ, Jackson NJ, Wieland W, Sun Y, Schwab RJ. An examination of methodological paradigms for calculating upper airway critical pressures during sleep. SLEEP 2016;39(5):977–987.

Significance.

Observed Pcrit values provide a consistent metric for describing hypotonic upper airway collapsibility that can be used as a reliable alternative to extrapolated Pcrit. As we found that observed Pcrit was more easily derived than extrapolated Pcrit, use of observed Pcrit values may make it easier to include measurements of upper airway critical closing pressure in clinical research related to sleep apnea.

INTRODUCTION

Obstructive sleep apnea (OSA) is a common disorder characterized by intermittent closure of the upper airway during sleep. To understand the mechanisms for dynamic collapse of the upper airway during respiration, its behavior is frequently modeled as a type of collapsible tube known as a Starling resistor.¹ Specifically, the Starling resistor is characterized by flow that initially increases linearly in response to increases in driving pressure. Above a critical upstream driving pressure, however, flow diminishes and ultimately plateaus, despite continued increases in driving pressure. In the upper airway, this characteristic behavior of Starling resistors is recognized as flow limitation, and the driving pressure at which complete collapse of the airway occurs with zero flow is known as the upper airway critical closing pressure (Pcrit).¹

Measurements of Pcrit are frequently used to characterize upper airway collapsibility, especially for research purposes.^2–5 Pcrit is determined by varying exogenously delivered upstream nasal pressure (P_N), measuring the resultant airflow, and extrapolating the pressure at which the upper airway occludes (i.e., Pcrit) from these data. Using this technique, the upper airway of individuals with OSA has been shown to be more collapsible (i.e., to have more positive Pcrit values) during sleep than those of normal individuals and snorers.² However, Pcrit measurements have generally been considered difficult to perform because of: (1) the technical expertise required to implement the protocols; (2) the labor-intensive nature of making the pressure-flow measurements; (3) the intricacies of examining the ensuing tracings for flow limitation; and (4) the relatively high proportion of studies that fail to generate usable data due to subjects failing to remain asleep during testing. Furthermore, different approaches for testing pressure-flow relationships over multiple nasal pressure levels and for analyzing these data to determine Pcrit exist.^4,6–9 These challenges have made it difficult to routinely use Pcrit in research or clinical practice and to compare data between different research studies.

In order to address these challenges, we decided to study different Pcrit analysis paradigms. The overall goal of this study was to examine whether simplifying and standardizing procedures for determining Pcrit could increase the proportion of subjects with usable data while providing reliable Pcrit values. Several methodological issues with important implications for the measurement of Pcrit were examined, specifically: (1) the utility of determining the Pcrit value from direct observation of occluded (no flow) breaths,^9,10 which we hypothesized could reliably describe upper airway collapsibility, versus inferring Pcrit from extrapolated data; (2) the use of Pcrit values generated by aggregating pressure-flow data over multiple individual runs of P_N levels versus Pcrit values averaged across individual pressure-flow runs, which we hypothesized would not be significantly different; (3) the utility of generating Pcrit values using only flow-limited breaths compared to values generated using nonflow-limited and flow-limited breaths, which we also hypothesized would not be significantly different; and (4) the utility of restricting valid extrapolated Pcrit values to those within a limited range of either the pressure at which upper airway collapse was observed or of the lowest nasal pressure applied during a given run (when observed Pcrit was not achieved).⁷ Finally, we sought to determine the relationship between observed and extrapolated Pcrit measures and severity of sleep-disordered breathing. We hypothesized that Pcrit would be lower (more negative) among normal individuals compared to those with obstructive sleep apnea, as has been previously reported in studies examining measurements of upper airway collapsibility.^2,11 Some of these results have been previously reported in abstract form.¹²

METHODS

Study Population

The study population included 351 adults with and without sleep-disordered breathing who were participants in 3 NIH-sponsored studies at the University of Pennsylvania Clinical Research Center for Sleep: (1) “Mechanisms of Menopausal Obstructive Sleep Apnea Study” (MeMOSA: n = 117), focused on midlife women with and without OSA^13,14; and two studies examining obese individuals with OSA and body mass index-matched controls before and after weight loss, (2) “Obesity & OSA: Understanding the Importance of Tongue Fat and Metabolic Function” (OBEY: n = 119),^15,16 and (3) “Understanding the Relationship between Obesity and Tongue Fat” (TOBE: n = 115).¹⁵ The apnea-hypopnea index (AHI) was determined using full overnight polysomnography in the laboratory or home,^14–17 as per the protocol of the contributing study. Hypopneas were scored in association with either ≥ 3% oxyhemoglobin desaturation or arousal.¹⁸ Participants with OSA were defined by an AHI ≥ 10 events/h and those without OSA by an AHI < 10 events/h. The studies were approved by the University of Pennsylvania Institutional Review Board. All participants provided written informed consent.

Pcrit Experimental Protocol

Pressure-flow relationships were measured during an in-laboratory polysomnogram in all subjects. The Pcrit experimental protocol was standardized across the three individual studies so the data could be pooled for prospective examination of different methodologic techniques for determining Pcrit. Routine polysomnographic measurements including electroencephalogram (C3/A2, C4/A1, Oz/A2), right and left electrooculogram, sub-mental electromyogram, electrocardiogram, chest and abdominal respiratory inductance plethysmography (Protech, Woodinville, WA) and pulse oximetry (Nellcor, Covidien-Nellcor, Boulder, CO) were obtained.¹⁹ Participants wore a nasal continuous positive airway pressure (CPAP) mask (Profile Lite, Philips Respironics, Murrysville, PA). Airflow was measured using a heated pneumotachometer (Hans Rudolph, Inc, Kansas City, MO) with a pressure transducer (AD Instruments, Colorado Springs, CO) attached to the CPAP mask. Nasal pressure was measured at the mask using a differential pressure transducer (Validyne Engineering Corp, Northridge, CA) referenced to atmosphere.

Signals were acquired on a PowerLab system (AD Instruments, Colorado Springs, CO) and simultaneously displayed on a Sandman polysomnography system (Natus Medical Embla, San Carlos, CA). Nasal pressure (P_N) was administered over a range of positive and subatmospheric pressures using a device provided by Philips Respironics (Murrysville, PA). The machine was calibrated from −20 to +20 cm H₂O prior to the studies and a toggle switch allowed participants to be switched rapidly between different nasal pressures. Actual pressures administered across all participants fell between these calibrated limits (−20 to +18 cm H₂O).

After sleep onset, the holding pressure, i.e., the P_N just above the pressure at which flow limitation was first discernible, was determined. Flattening of the pressure-flow curve (not reduced flow amplitude) was used to identify flow limitation and was used as a surrogate for true inspiratory flow limitation (the flow-time contour).^20,21 For participants without evidence of flow limitation when P_N = 0 cm H 2O, a holding pressure of +1 cm H₂O was used. After 5 min of stable stage N2 and/ or stage N3 sleep, P_N was decreased rapidly from the holding pressure by 2 cm H₂O for five breaths, and then returned to the holding pressure for 1 minute. This rapid decrease in P_N was then repeated, with 1 to 2 cm H₂O increment increases in the magnitude of the drop in P_N with each repetition to achieve progressively lower pressures, and a return after each P_N drop to the holding pressure until either arousal from sleep or an obstructive apnea occurred (Figure 1). If an arousal occurred, the protocol was resumed after the individual reinitiated stable stage 2/3 sleep. Each series of pressure drops was labeled an individual “run.” This hypotonic technique has been shown to result in an upper airway with minimal neuromotor activation.²² A median (range) of 5 (1–14) runs using this technique were performed per participant to obtain a range of P_N and peak inspira-tory airflow (VImax) values during airflow limitation. For each participant, attempts were made to complete three runs during which airway closure (i.e., an obstructive apnea) was achieved.

Figure 1.

Hypotonic Pcrit protocol. This schematic illustrates the protocol for obtaining hypotonic Pcrit (red line), with distinct stepwise reductions in nasal pressure for five breaths, and return to holding pressure in between each reduction in nasal pressure. Progressive reductions in airflow are seen with greater reductions in nasal pressure (blue line), culminating in an apnea. Artifactually high and low flow values observed during the first breath with changes in nasal pressure were excluded from analysis.

Pcrit Data Analysis

Using guidance from the existing literature^6–8,20 and the investigators' prior experience, scoring rules were developed to provide consistent, objective criteria for determining which data to include when calculating Pcrit and to allow us to perform a rigorous comparison of different analytic techniques for measuring Pcrit, as described in the next paragraphs.

Calculation of Peak Inspiratory Flow Values

The mean peak inspiratory flow of the holding pressure breaths was calculated using six non-flow-limited breaths recorded just prior to the start of each run (Figure 2A). For individual breaths within individual runs, the peak inspiratory flow (Figure 2B) was determined by subtracting the mean of the total inspiratory and expiratory flow from the maximum inspiratory flow value.⁸

Figure 2.

Airflow tracings. (A) Airflow tracing before and during a drop in holding pressure from a subject in stable NREM (stage N2) sleep. At left, PN is dropped from the holding pressure for five breaths, inducing flow limited breathing (right), before being set back to the holding pressure. Mean inspiratory and expiratory flow are determined from 6 breaths at holding pressure (solid blue line, left), and from breaths 2–4 (solid blue line, right) during the drop from holding pressure. (B) Single breath airflow tracing for peak airflow calculation. Mean flow for each individual breath was calculated by selecting the maximum inspiratory flow value for each breath and calculating the peak flow as the difference between the maximum and mean inspiratory flow values.

Identification of Flow-Limited, Nonflow-Limited, and No-Flow Breaths

Each run of pressure-flow curves was reviewed by a trained technician, who identified flow-limited breaths using the presence of at least one of the following criteria: a flattened inspiratory flow contour; negative effort dependence (decrease in flow with corresponding increase in effort); asynchrony (differential in timing of respiratory effort and flow between breaths); or snoring (high frequency oscillations in pneumotach sensor).^20,21,23,24 No-flow breaths were identified as breaths with a peak flow less than 50 mL/sec.⁸ Breaths not meeting any of these criteria were considered nonflow-limited breaths.

Determination of Valid Breaths within Pressure Levels

To determine whether a breath was valid for inclusion in the analyses, several criteria were examined. First, breaths were required to be from supine stage 2 or stage 3 sleep. The flow signal was examined visually and quantitatively for mask leak and breaths with excessive leak causing discrepancies between the machine pressure and mask pressure were excluded. In the primary analyses (details in the following paragraphs), only breaths that had been visually scored as flow-limited or no-flow were included. The analyses were next restricted to breaths 2 to 4 from each set of five breaths at a given nasal pressure (Figure 2A).⁸ Breath 1 was excluded due to artifact created by the pressure drop and breath 5 was excluded due to occasional arousals or artifact created by the abrupt return to the holding pressure. Finally, breaths with a peak flow greater than the maximum peak inspiratory flow observed at the holding pressure and breaths that occurred during or subsequent to an arousal were excluded.²⁵

Determination of Valid Pressure Levels within Runs

After applying the aforementioned filtering criteria to the individual breaths, the number of valid flow-limited or no-flow breaths at each nasal pressure level within each Pcrit run was examined. For breaths at a given P_N to be included in further analyses, at least two valid flow-limited breaths or one no-flow breath were required. All breaths from nasal pressure levels lower than the highest nasal pressure at which no-flow breaths were first observed were also excluded.

Determination of Valid Runs for Extrapolation

Next, we determined whether each run included adequate data for calculating an extrapolated Pcrit value by fitting a linear regression model. The number of valid nasal pressure levels within each run was examined. Runs included in extrapolated analyses were required to have a minimum of three nasal pressure levels with valid flow-limited breaths, or a minimum of two nasal pressure levels when no-flow breaths were observed during the run.

Calculation of Pcrit Measurements

Next, observed and extrapolated Pcrit values were calculated for each study participant.

Mean across Individual Run and Aggregate Data Methodologies for Extrapolated Pcrit

For each Pcrit measure described in the next paragraphs, subject-specific values were computed using two methodologies: (1) Pcrit values were calculated separately for each individual run and then averaged (‘mean across individual runs’ technique); and (2) data from individual runs were combined into one plot and a single Pcrit value was determined using all valid data (‘aggregated data’ technique). Both types of values were generated to determine whether statistically significant differences existed between Pcrit values derived from the mean of individual runs, which we believed captured modest differences between runs, and aggregate Pcrit. Figure 3 illustrates the two techniques.

Figure 3.

Schematic illustration of individual run and aggregated Pcrit methods. A hypothetical example of three individual Pcrit runs and the associated regression lines for determining extrapolated Pcrit (shown in blue, green, and red), as well as the aggregate regression line (black dashed line) derived by combining the data across the individual runs. The three individual runs result in extrapolated Pcrit estimates (at 0 ml/ sec) of −0.71 (blue), −4.29 (green) and −2.32 (red) cm H2O, and a mean across individual runs of −2.43 cm H2O. Deriving the extrapolated Pcrit value by aggregating all of the data points across individual runs results in an estimate of −3.33 cm H2O (black dashed line).

Determination of Observed Pcrit Values

The observed Pcrit pressure level (P_N^OBS) was defined as the first P_N at which a participant had at least two breaths with peak inspiratory flow < 50 mL/sec, a threshold that has been validated previously.⁸ The observed Pcrit value (Pcrit_OBS) was then calculated as the mean nasal pressure level at the CPAP mask of all breaths with peak inspiratory flow < 50 mL/sec within the identified P_N^OBS. Because slight variations were noted in pressures measured at the mask between breaths due to fluctuations in inspiratory flow, actual measured mask pressures rather than machine-delivered pressures were used for calculating Pcrit.

Secondary analyses were performed to examine the effect of defining the observed Pcrit value as the first P_N at which a subject had at least one breath (instead of two) with peak inspiratory flow < 50 mL/sec (see supplemental material).

Determination of Extrapolated Pcrit Values

To calculate an extrapolated Pcrit value for each participant, a linear regression model was fitted to all valid data points¹ (Figure 3). Specifically, peak inspiratory flow (y-axis) for all valid breaths was regressed against the nasal pressure at the CPAP mask (x-axis). Then, a raw extrapolated Pcrit value (Pcrit_EX^RAW) was calculated as the predicted nasal pressure at which airflow was equal to 0 ml/sec (i.e., the x-intercept of the linear regression line). Values of Pcrit_EX^RAW falling outside of the calibrated pressure limits (i.e., −20 to +20 cm H₂O) were set equal to the closest pressure limit (e.g., a Pcrit_EX^RAW of −30 would be set equal to −20 cm H₂O). Based on a previous publication,⁵ a filtered extrapolated Pcrit value excluding any Pcrit_E XRAW values more than 3 cm H₂O higher or lower than either the first P_N at which a participant had at least two breaths with peak inspiratory flow < 50 mL/sec (i.e., the observed Pcrit pressure) or the lowest nasal pressure with valid breaths (when the noflow state was not achieved) was created.

Extrapolated Pcrit Calculation Including a Subset of Nonflow-Limited Breaths

To determine how the exclusion of nonflow-limited breaths affected the proportion of participants for whom Pcrit could be extrapolated and the effect on average Pcrit values, secondary analyses were performed in which nonflow-limited breaths with peak flow values below the maximum peak inspiratory flow observed at the holding pressure were retained. Extrapolated Pcrit was calculated using the expanded dataset (nonflow-limited, flow-limited, and no-flow breaths) while applying the other criteria as specified.

Extrapolated Pcrit Calculation to 50 mL/sec

Extrapolated Pcrit values were recomputed as the predicted nasal pressure at which airflow was equal to 50 mL/sec in the linear regression model, rather than 0 mL/sec. This definition provides a measurement more closely related to the study definition of observed Pcrit, i.e., the mean nasal pressure level of all breaths at the first P_N with peak inspiratory flow < 50 mL/sec.

Statistical Analyses

Continuous variables were summarized using means, standard deviations (SD), and ranges; categorical variables were summarized using frequencies and percentages. In order to compare the utility of the various techniques for determining passive Pcrit in the absence of a “gold standard” analytic technique, we examined differences between the techniques in two ways: by using linear mixed models to compare least squares means and differences and 95% confidence intervals (CI) between pairs of Pcrit values; and by generating Bland-Altman plots to examine the extent of agreement between different techniques. Specifically, pairwise comparisons of Pcrit values generated using the mean of individual runs and aggregated data were performed, using both extrapolated and observed methods. Because we could not calculate all Pcrit values for all participants, and to control for within-subject correlation when both measures were available for the same participant, linear mixed models were used to compare different Pcrit values among subjects. Models were fit using a random subject effect, with the calculated Pcrit value as the outcome and a categorical term indicating the technique used (e.g., mean individual run Pcrit_EX^±3 versus aggregated Pcrit_EX^±3). Results from mixed models are reported as adjusted least squares mean difference (Δ) and 95% CI. Pcrit analytic techniques demonstrating consistently less variability (as evidenced by the SD) and smaller mean differences and 95% CI in pairwise comparisons were assessed to be more reliable than analytic techniques producing larger variability and mean differences and 95% CI.

In addition to estimating the mean difference (or bias) between different techniques using linear mixed models, methods described by Bland and Altman^26,27 were utilized to examine agreement between Pcrit values derived using different techniques within the subset of subjects with valid measurements for both techniques. For each pair of Pcrit measures compared (n = 60 total, see details in the supplemental material), the subject-specific difference in measurements was examined against the average of the two measures (i.e., the “Bland-Altman plot”) for systematic bias, and the mean difference within the sample and the limits of agreement (mean difference ± 2 SD) were calculated. Based on previous work,⁷ limits of agreement of more than ± 3 cm H₂O were considered a clinically signifi-cant difference in the two techniques. Techniques consistently demonstrating Pcrit values within the limits of agreement and with less variability (i.e., smaller SD) were assessed to be of greater utility than techniques generating values consistently falling outside the limits of agreement when compared with other techniques, and with greater variability.

The relationship between Pcrit measures and OSA was examined in two ways. Unpaired t-tests were used to compare mean Pcrit values between participants with OSA (AHI ≥ 10 events/h) and without OSA (AHI < 10 events/h).^15,16 The association between Pcrit values and continuous AHI was also assessed using Pearson correlation coefficients; for these analyses, AHI was natural log-transformed for normality.

This exploration into different methodologies for the measurement of Pcrit resulted in several a priori hypotheses, each requiring multiple relevant comparisons. To control for these multiple comparisons, Bonferroni corrected P thresholds ranging from 0.025–0.0025 were used to determine statistical significance for each hypothesis (see detailed hypotheses, number of comparisons and individual Bonferroni-corrected levels of significance in supplemental material). However, given the exploratory nature of these analyses, any value of P < 0.05 was considered suggestive evidence of a statistically significant difference. In addition to statistical significance, results were interpreted with regard to whether the observed magnitude of the difference was clinically significant.

RESULTS

Demographic Characteristics of Study Sample

There were 351 participants, including 106 men (30%) and 245 women (70%) (Table 1). Pcrit was able to be determined using at least one of the study definitions in 308 individuals (88%). Among the 43 individuals (12%) among whom Pcrit could not be determined, Pcrit determination most frequently failed (n = 29) because they did not sleep in the supine position during any Pcrit runs. Pcrit determination failed in other participants (n = 14) due to a combination of factors, including lack of supine sleep in some Pcrit runs and/or ≤ 2 valid pressure levels with flow-limited breaths after applying the various filtering criteria detailed in the Methods.

Table 1.

Demographic characteristics of the study sample.

Observed Pcrit Values

Average Pcrit_OBS values based on the mean across individual runs and aggregate data techniques are presented in Table 2 and Table S1 (in supplemental material).

Table 2.

Observed Pcrit values (≥ 2 breaths), overall and for obstructive sleep apnea and no obstructive sleep apnea groups.

When P_N^OBS was defined as the first pressure level with ≥ 2 breaths with peak flow < 50 mL/sec, observed Pcrit values were calculated in 234 participants (67%) using the mean across individual runs and 242 participants (69%) based on aggregated data (Table 2). Pcrit_OBS based on aggregated data was significantly less negative than Pcrit_OBS calculated as the mean across individual runs in mixed model analysis (Δ [95% CI]: 1.01 [0.87, 1.15]; P < 0.0001), and the limits of agreement extended beyond the clinically significant threshold of ± 3 cm H₂O (Table S3, supplemental material).

Pcrit Values Extrapolated to 0 mL/sec

Average values for Pcrit extrapolated to 0 mL/sec from the mean across individual runs and from aggregate data are presented in Table 3. Values are reported separately for Pcrit extrapolated entirely from flow-limited and no-flow breaths, and for Pcrit values extrapolated from data including nonflow-limited breaths.

Table 3.

Extrapolated Pcrit values, overall and for obstructive sleep apnea and no obstructive sleep apnea groups.

When data were restricted to flow-limited and no-flow breaths, Pcrit_EX^RAW values were calculated for 237 participants (68%) using the mean across individual runs and 227 participants (65%) based on aggregated data (Table 3). There was no statistically significant difference between the Pcrit_EX^RAW values based on the technique used (Δ [95% CI]: −0.15 [−1.26, 0.96], P = 0.794). However, in the Bland-Altman analyses of agreement, pairwise comparisons involving any of the Pcrit EXRAW values resulted in wide limits of agreement, all with bounds greater than ± 3 cm H₂O (Table S3).

After applying the filtering criterion stipulating that extrapolated values must be within 3 cm H₂O of the lowest valid nasal pressure level (shaded data in Table 3), there was a 13% decrease in the number of subjects with valid measures using the mean across individual runs (n = 206 [59%]). When the filtering criterion was applied to the aggregated data, the number of participants with valid measures was reduced by 45% (n = 124 [35%]). Using linear mixed models, there was no significant difference in mean Pcrit_EX^±3 values between these two techniques (Δ [95% CI]: 0.03 [−0.23, 0.28]; P = 0.837). Furthermore, the Bland-Altman analyses of agreement detected no significant difference (or bias) between filtered measures using flow-limited and no-flow breaths only, and limits of agreement of ± 2.7 cm H₂O when comparing agreement between Pcrit_EX^±3 values based on the mean across individual runs or the aggregate data technique and extrapolated to 0 mL/sec (Table S3).

Including nonflow-limited breaths in the dataset increased the number of participants for whom Pcrit could be extrapolated: Pcrit_EX^RAW values were calculated in 298 participants (85%) using the mean across individual runs technique and 299 participants (85%) when based on aggregated data (Table 3). Although mean values between the two techniques were not significantly different using linear mixed modeling (Δ [95% CI]: −0.60 [−1.49, 0.28]; P = 0.179), Bland-Altman limits of agreement extending beyond ± 3 cm H₂O were observed (Table S3). For Pcrit_EX^±3, although more participants were retained by including nonflow-limited breaths, the proportion of individuals with values meeting the filtering criteria fell sharply for Pcrit values extrapolated from aggregate data (as occurred when the dataset included only flow-limited and no-flow breaths): values were calculated for 245 subjects (70%) using the mean across individual runs and 135 subjects (38%) based on aggregated data. The Pcrit_EX^±3 value using the aggregate data was slightly more negative compared to the Pcrit_EX^±3 value using the mean across individual runs, but this difference was not statistically significant using linear mixed modeling (Δ [95% CI]: −0.16 [−0.35, −0.02]; P = 0.074). Likewise, although the Bland-Altman analysis found the difference between the measures to be of borderline statistical significance (P = 0.031), limits of agreement were observed to be within the definition of clinical nonsignificance (−2.29 to 1.89 cm H₂O; Table S3).

Including nonflow-limited breaths in the dataset produced lower mean Pcrit_EX^RAW and similar Pcrit_EX^±3 values compared to data restricted to flow-limited and no-flow breaths (Table 3). None of these comparisons met the specified Bonferroni corrected level of significance (P < 0.00625; Table S4, supplemental material). Comparisons of Pcrit_EX^RAW using Bland-Altman analyses demonstrated limits of agreement extending beyond ± 3 cm H₂O. However, when comparing Pcrit_EX^±3 values using flow limited and no-flow breaths to values using all breaths within the individual run and aggregated techniques, limits of agreement within ± 3 cm H₂O were observed (Table S4).

Pcrit Values Extrapolated to 50 mL/sec

To provide an extrapolated Pcrit measurement more consistent with the observed Pcrit definition (i.e., breaths with peak inspiratory flow < 50 mL/sec), Pcrit_EX^RAW and Pcrit_EX^±3 values were recomputed as the pressure at which peak flow = 50 mL/sec (Table 3).

Because Pcrit extrapolated to 50 mL/sec is derived from the identical regression line as Pcrit extrapolated to 0 mL/ sec, Pcrit_EX^RAW values were available for the same subsets described in the previous section. As expected, Pcrit measurements extrapolated to 50 mL/sec were significantly more positive compared to corresponding Pcrit values extrapolated to 0 mL/sec (all P < 0.0001). When comparing Pcrit values across techniques using flow-limited and no-flow breaths only, there was no significant difference (Δ [95% CI]: 0.03 [−0.89, 0.95]; P = 0.952) in Pcrit_EX^RAW between the aggregate data and mean across individual runs. Similarly, there was no significant difference in the Pcrit_EX^RAW between the 2 techniques when including nonflow-limited breaths (Δ [95% CI]: −0.43 [−1.28, 0.42]; P = 0.318).

When using this higher threshold for extrapolated Pcrit using flow-limited and no-flow breaths, filtering values based on the 3 cm H₂O rule resulted in less exclusion (Table 3). Mean values were not significantly different between the mean across individual runs or aggregated data techniques, based on a linear mixed model (Δ [95% CI]: −0.03 [−0.21, 0.16]; P = 0.774).

Similar results were observed for Pcrit_EX^±3 when nonflow-limited breaths were included in the dataset (Table 3). Pcrit measurements based on the mean across individual runs were significantly higher compared to those using aggregated data (Δ [95% CI]: 0.35 [0.20, 0.51]; P < 0.0001).

Pcrit_EX^RAW values extrapolated to 50 mL/sec that included all breaths were compared to those restricting the dataset to flow-limited and no-flow breaths. Values including all breaths were significantly lower whether the mean was determined across individual runs (Δ [95% CI]: −1.04 [−1.89, −0.19]; P = 0.017) or the aggregated data (Δ [95% CI]: −1.63 [−2.55, −0.71]; P = 0.001). However, only the difference in aggregated Pcrit values was statistically significant after correction for multiple comparisons. Similarly, Pcrit_EX^±3 values from data including nonflow-limited breaths were significantly lower only when using aggregated data (Δ [95% CI]: −0.42 [−0.57, −0.27]; P < 0.0001). Results from Bland-Altman analyses for comparisons of Pcrit measurements extrapolated to 50 mL/sec (Tables S3 and S4) were similar to those for Pcrit extrapolated to 0 mL/sec, i.e. demonstrating clinically significant limits of agreement between comparisons of Pcrit_EX^RAW, and limits of agreement that were not clinically significant between comparisons of Pcrit_EX^±3.

Comparisons between Observed and Extrapolated Pcrit Values

Observed Pcrit measurements were generally more positive compared to Pcrit values extrapolated to 0 mL/sec flow (Tables 2 and 3), whether using the mean across individual runs or aggregated data techniques. However, because Pcrit_OBS was defined as the average nasal pressure at the first pressure with peak flow < 50 mL/sec, this was expected. When compared to Pcrit values extrapolated to 50 mL/sec, observed Pcrit values still tended to be greater. However, differences were smaller and agreement was better between observed Pcrit and Pcrit extrapolated to 50 mL/sec than between observed Pcrit and Pcrit extrapolated to 0 mL/sec.

Mean differences between observed Pcrit using the ≥ 2 breath P_N^OBS definition and corresponding Pcrit values extrapolated to 0 mL/sec ranged from 2.0 to 4.3 for Pcrit_EX^RAW values and from 1.0 to 1.7 cm H₂O for Pcrit_EX^±3 (all P < 0.0001 for both linear mixed models and paired comparisons). The Bland-Altman limits of agreement comparing ≥ 2 breath Pcrit_OBS and extrapolated Pcrit at 0 mL/sec exceeded the threshold for a clinically significant difference for all comparisons except those between the individual run mean Pcrit_EX^±3 values and Pcrit extrapolated to 0 mL/sec (Table S6, supplemental material).

When comparing ≥ 2 breath Pcrit_OBS to extrapolated values at 50 mL/sec, as opposed to 0 mL/sec, better levels of agreement were observed between measures (Table S6). Pcrit_OBS was significantly greater (all P < 0.0001) than extrapolated Pcrit to 50 mL/sec using the aggregated data technique, with differences ranging from 0.5 to 2.7 cm H₂O, and limits of agreement extended beyond ± 3 cm H₂O for all but the Pcrit_EX^±3 value using flow-limited and no-flow breaths. Conversely, when comparing Pcrit_OBS to extrapolated Pcrit values using the mean across individual runs, a statistically significant difference was observed only between ≥ 2 breath Pcrit_OBS and Pcrit_EX^RAW using all breaths (P < 0.0001). However, the Bland-Altman limits of agreement extended beyond the bounds of clinical significance for both comparisons of Pcrit_OBS and Pcrit_EX^RAW at 50 mL/sec using the mean across individual runs (Table S6). The best agreement was seen between the ≥ 2 breath Pcrit_OBS and Pcrit_EX^±3 based on the mean across individual runs, whether all breaths or only flow-limited and no-flow breaths were used. Importantly, using Bland-Altman methods to examine the paired data, good agreement between these two techniques was observed, with limits of agreement of approximately ± 2 cm H₂O (Table S6).

Relationship between Pcrit Values and AHI

The relationship between various Pcrit measurements and OSA was examined in two ways: (1) by comparing mean Pcrit values between normal participants (AHI < 10) and participants with OSA (AHI ≥ 10 events/h) (Tables 2 and 3); and (2) by examining the Pearson correlation between Pcrit measurements and natural log transformed AHI (Table 4).

Table 4.

Pearson correlations between Pcrit values and natural log-transformed apnea-hypopnea index.

Comparisons of observed Pcrit measures between individuals with and without OSA are presented in Table 2. For each Pcrit value, observed mean Pcrit values in those with OSA were more positive than observed Pcrit values in the normal group (differences in mean values ranging from 2.7 to 2.8 cm H₂O; P < 0.0001). For each Pcrit_OBS value, mean Pcrit_OBS was > 0 cm H₂O among the group with OSA, indicating that the airway closed above atmospheric pressure, whereas values among those without OSA were < 0 cm H₂O.

Comparisons of extrapolated Pcrit measures in participants with and without OSA are presented in Table 3. Mean Pcrit values among those with OSA were again more positive compared to those without OSA (mean differences ranging from 2.4 to 2.8 cm H₂O). All P values for comparisons of Pcrit_EX^±3 between groups with and without OSA met the specified Bonferroni corrected level of significance (P < 0.0025) and were associated with large effect sizes (ranging from 0.91 to 1.24). In contrast, comparisons of unfiltered extrapolated Pcrit were generally not statistically significant after correction for multiple comparisons, reflecting the larger variability in Pcrit_EX^RAW.

Correlation analyses examining the amount of shared variability between observed and extrapolated Pcrit values and natural log-transformed AHI are presented in Table 4. Signifi-cant positive correlations were observed between all Pcrit measures and AHI (all P < 0.0001). However, correlations between AHI and Pcrit_EX^RAW were only about half as large as those for Pcrit_OBS or Pcrit_EX^±3 (see shaded data in Table 4). The results suggest that Pcrit_OBS or Pcrit_EX^±3 explains a considerable proportion of the variability in AHI (25% variability explained, i.e., a correlation coefficient of approximately 0.5), whereas Pcrit_EX^RAW only accounts for 7.4% of the variability on average.

DISCUSSION

The objective of this study was to examine different paradigms for analyzing upper airway pressure-flow data and determining critical closing pressures during sleep. Given the potentially powerful physiological justification to be made for direct observation of the pressure at which the upper airway collapses, we were especially interested in examining whether there were significant differences between observed Pcrit, which has been infrequently reported in determinations of upper airway collapsibility, and extrapolated Pcrit. Although at least one measure of Pcrit was available in 88% of the sample, this number varied depending on the method (observed versus extrapolated, mean of individual runs vs. aggregate) and technique (use of one or two breaths to define observed Pcrit, extrapolation to 0 or 50 mL/sec, inclusion or exclusion of nonflow-limited breaths, application of ± 3 cm H₂O filter). As expected, raw (unfiltered) extrapolated Pcrit values tended to have fewer missing values; however, they also had the greatest variability. Our results demonstrate that observed Pcrit values provided a consistent metric for describing hypotonic upper airway collapsibility that can be used as a reliable alternative to extrapolated Pcrit. Among extrapolated Pcrit measurements, application of a series of filtering criteria to optimize the validity of the data likewise generated Pcrit values that differentiated between groups with and without obstructive sleep apnea.

Observed Pcrit Values

Observed Pcrit was able to be determined in 67% to 69% of participants, a similar or higher proportion of subjects than those who met all criteria for determination of extrapolated Pcrit after applying the ± 3 cm H₂O filter. Observed Pcrit values from the aggregated data were significantly more positive compared to the values from the mean across individual runs. Given the methodology used to define P_N^OBS, this observation was expected. P_N^OBS was defined as the first pressure level where a subject had breaths with peak inspiratory flow < 50 mL/sec. Thus, aggregated Pcrit_OBS (equal to the mean pressure of all breaths with flow < 50 mL/sec at P_N^OBS) should always be equal to or greater than the maximum Pcrit_OBS across individual runs, because deriving Pcrit from the aggregated data essentially yields the maximum of the values from individual runs.

Pcrit Values Extrapolated to 0 mL/sec

Significant differences were generally not observed between Pcrit values extrapolated to 0 mL/sec using the various techniques, regardless of whether raw or filtered values were used, the dataset included all breaths or was restricted to flow-limited and no-flow breaths, or the comparisons were within or between aggregated or individual run techniques. In part, this is likely due to the exclusion earlier in the analytic process of raw peak inspiratory flow values greater than the maximum peak inspiratory flow observed at the holding pressure and breaths during or subsequent to an arousal, as described in Methods. Differentiating between nonflow-limited, flow-limited, and noflow breaths recorded during nasal pressure drops in the hypo-tonic Pcrit protocol is a labor-intensive task requiring training to reliably apply the multiple different criteria used to identify flow-limited breathing. The study results suggest that making these distinctions may not be needed to generate most extrapolated Pcrit values, assuming that data are filtered in the way we describe. In contrast, raw mean Pcrit values had large standard deviations and comparisons of these values yielded wide Bland-Altman limits of agreement. Thus, raw extrapolated values resulted in unreliable and highly variable Pcrit pressure values.

Overall, the proportion of participants with a valid Pcrit_EX^±3 based on aggregated data was markedly smaller than the proportion with a valid Pcrit_EX^±3 using the mean across individual runs technique. This is likely to be due to the increased variability associated with extrapolating data from multiple runs included in a single plot. When a raw Pcrit value more than ± 3 cm H₂O higher or lower than the lowest valid nasal pressure applied during the study was generated from an individual run and thus excluded, participants typically had other valid runs that were retained so that Pcrit could be successfully calculated. However, because only a single Pcrit value was generated using the aggregate method, values outside the ± 3 cm H 2O range led to the exclusion of all data for a given subject. When Pcrit_EX^±3 was compared among subjects with valid data for both the mean of individual runs and aggregate data techniques, some comparisons achieved statistical significance. However, absolute differences were small (all < 0.5 cm H₂O) and not clinically significant.

The proportion of subjects among whom Pcrit could be calculated was higher for Pcrit_EX^RAW compared to Pcrit_EX^±3. However, Pcrit_EX^RAW (unfiltered) values were characterized by considerably more variability (i.e., less precision) than Pcrit_EX^±3 values. Thus, we recommend use of the ± 3 cm H₂O filtering criterion when calculating extrapolated Pcrit, as has been utilized in previous studies.⁵ Furthermore, because Pcrit was quantifiable in a larger proportion of the sample using the individual run data, we recommend that extrapolated Pcrit be determined using the mean Pcrit across individual runs, rather than aggregated data. The loss of sample resulting from applying the ± 3 cm H₂O filtering criterion can be mitigated by including nonflow-limited breaths in the dataset, as absolute differences from mean Pcrit values from flow-limited and no-flow breaths alone were small (all < 0.5 cm H₂O) and narrow Bland-Altman limits of agreement were observed, suggesting the absence of a clinically or statistically significant difference between the techniques.

Pcrit Values Extrapolated to 50 mL/sec

As expected, Pcrit values extrapolated to 50 mL/sec were significantly higher than values extrapolated to 0 mL/sec. Fifty mL/sec was chosen as the secondary extrapolation point given the definition of no-flow breaths as those with peak inspiratory flow of < 50 mL/sec, a definition used in this study and others.⁸ Pcrit_EX^RAW (unfiltered) values extrapolated to 50 mL/sec had large standard deviations, similar to Pcrit^RAW extrapolated to 0 cm/sec, and their use is not recommended.

When the ± 3 cm H₂O filter was applied to raw Pcrit values extrapolated to 50 mL/sec, more subjects were retained in the sample compared to extrapolation to 0 mL/sec. This was because Pcrit values extrapolated to 50 ml/sec were higher and thus more likely to fall within 3 cm H₂O of the lowest nasal pressure with valid breaths (typically, observed Pcrit). Loss of samples was again further mitigated by including nonflow-limited breaths. The Bland-Altman limits of agreement were within the range of clinical nonsignificance when comparing all breaths to only flow-limited and no-flow breaths, suggesting a lack of clinically or statistically significant differences between these techniques. Given these findings, we recommend that Pcrit be extrapolated to 50 mL/sec (as opposed to 0 mL/sec), which generally agrees with expert visual analysis of no-flow breaths and maximizes the proportion of subjects with usable data.

Comparisons between Observed and Extrapolated Pcrit Values

While observed Pcrit values using ≥ 2 breaths were consistently higher than Pcrit_EX^±3 values extrapolated to 0 mL/sec, there were fewer significant differences between ≥ 2 breath observed Pcrit values and Pcrit_EX^±3 when extrapolated to 50 mL/ sec. This finding is consistent with the use of a boundary of peak flow < 50 mL/sec for breaths used to determine observed Pcrit. Furthermore, it indicates that although Pcrit values extrapolated to 50 mL/sec better estimated observed Pcrit in these data, observed Pcrit provides a reliable alternative descriptor of upper airway collapsibility to extrapolated Pcrit.

Relationship between Pcrit Values and the AHI

Both extrapolated and observed Pcrit values were higher among individuals with OSA compared to those without OSA, and these differences were highly significant. Thus, study methods for determining extrapolated and observed Pcrit were able to distinguish between subjects with OSA and normal individuals. Furthermore, mean observed Pcrit was > 0 cm H₂O among participants with OSA, indicating that the airway closed at atmospheric pressure. Among normal individuals, mean observed Pcrit was < 0 cm H₂O, indicating that active negative pressure was required to close the airway among those without OSA. These observations are consistent with previously reported extrapolated Pcrit values from other groups working in this area^2,28 and support the validity of our definitions.

Due to the greater variability in Pcrit_EX^RAW values, these measures resulted in less significant P values in comparisons of the OSA group and normal individuals when compared to results for Pcrit_OBS and Pcrit_EX^±3. Effect sizes using Pcrit_EX^RAW were two to three times smaller than those for filtered values. For Pcrit_OBS, effect sizes were similar across all four measures. Highly significant correlations between AHI and all Pcrit measures were observed. Moderate correlations were observed between AHI and Pcrit_OBS or Pcrit_EX^±3, with results suggesting that approximately 25% of the variability in AHI can be explained by hypotonic properties of the upper airway as captured using Pcrit_OBS or Pcrit_EX^±3. These findings are consistent with prior studies^22,29–31 demonstrating that, in addition to the passive propensity of the upper airway for collapse, other mechanisms such as upper airway muscle dilator activity contribute to the frequency and severity of apnea and hypopnea events. Correlations between Pcrit_EX^RAW and AHI were noticeably smaller. For this reason and others previously noted, the use of Pcrit_OBS and Pcrit_EX^±3 values is favored over the use of Pcrit_EX^RAW.

Strengths and Limitations

Particular strengths of this study include the large number of participants, which included individuals with and without OSA, and the inclusion of a large proportion of women, who have less frequently been included in studies of upper airway collapsibility. Previous studies examining Pcrit methodologic concerns have generally included much smaller numbers of mostly male subjects;^{1,2,6,8,11,22} Pcrit measurements have been performed in only a few larger studies.^7,32,33 The current study standardized Pcrit protocols among three different studies so that data could be collected prospectively and pooled for comparison of various methodologic analysis techniques. There are also several limitations. First, without a “gold standard” measurement of passive closing pressure, such as was performed by Isono et al. using direct observation of upper airway collapse under conditions of general anesthesia and muscle paralysis,³⁴ a true physiologic closing pressure is not available for comparison to values generated by the various analytic techniques. Next, esophageal manometry, which can improve the accuracy of identifying flow-limited breaths, was not utilized. However, previously published methods for reliably determining the presence of flow limitation based on visual inspection of flow waveforms^21,35,36 were utilized. In formulating rules for including individual runs in extrapolated analyses, a minimum of three valid nasal pressure levels were required when only flow-limited breaths were observed, but only two valid nasal pressures were required when no-flow breaths were observed, because datapoints at these pressures were less likely to produce high leverage outliers. However, extrapolated Pcrit values, especially after application of the ± 3 cm H₂O filter, were generally similar to the observed values for Pcrit, suggesting that using the lower minimum number of valid pressure levels when no-flow breaths were observed did not introduce undue variability to these values.

Other sources of variability in the Pcrit values exist, including true physiologic variability between runs, errors in identification of mean flow, and differences by sleep stage (i.e., stage 2 versus stage 3 sleep). Physiologic variability may be influenced by factors including neck and head position that are known to affect Pcrit,^4,37,38 that we were unable to measure and include in our analyses. Errors in identification of mean flow were minimized by having trained technicians extract data using standardized techniques. Several studies have failed to demonstrate significant differences in Pcrit between NREM sleep and REM sleep,^6,22 and between sleep stages N2 and N3,^39,40 during which measurements in the current study were made.

Full baseline polysomnography to define AHI was performed either at home or in the sleep laboratory, which could increase variability in AHI and lead to misclassification of AHI category among some subjects. However, such misclassification should bias toward the null, i.e., decrease the likelihood of detecting significant differences between Pcrit values in obtained in those with and without OSA. It should also be noted that Pcrit values in the group with OSA tended to be lower than reported in some earlier studies.^1,2 Although differences in measurement technique may account for such differences, another reason is likely to be the higher proportion of women in this study compared to other studies in which Pcrit was measured in OSA patients. Passive Pcrit has been observed to be lower in women compared to men after matching for body mass index, age, and disease severity,⁵ and in fact, we found lower Pcrit values in women compared to men using all techniques in the current study (mean differences ranging from −1.0 to −3.3 cm H₂O). Finally, determination of observed Pcrit values does not rely on a linear regression equation. Thus, upstream resistance (defined as the reciprocal of the slope, Pn/ ΔV_Imax cm H2O/L/s), which adds another dimension to measurements of upper airway collapsibility, cannot be calculated using the observed Pcrit technique.

The majority of participants in this study had known OSA or risk factors for increased upper airway collapsibility, such as obesity, postmenopausal status, or older age. Thus, these results may not be applicable to subjects with very low upper airway collapsibility, such as children, in whom Pcrit_OBS is less likely to be reached.

CONCLUSIONS

This exploration of different paradigms for determining passive Pcrit in a very large sample leads us to conclude that observed Pcrit values provide a reliable descriptor of hypotonic upper airway collapsibility compared to extrapolated Pcrit values. In addition, although the protocol for data acquisition is the same, observed Pcrit has the advantage of being directly witnessed and more easily derived than extrapolated Pcrit. Moreover, the observed method yielded reliable, usable Pcrit values among a similar or higher proportion of study subjects than extrapolated Pcrit values using the ± 3 cm H₂O filter.

Using multiple techniques for determining Pcrit allowed us to derive passive Pcrit among 88% of all participants who underwent the Pcrit protocol, including many naïve to CPAP therapy who may have been less tolerant of the experimental protocol. We thus propose that future protocols should reserve the calculation of extrapolated Pcrit to those who do not achieve the observed no-flow state. As observed Pcrit more closely approximated Pcrit extrapolated to 50 mL/sec than Pcrit extrapolated to 0 mL/sec, we further propose that for greatest utility, either observed Pcrit or Pcrit_EX^±3 extrapolated to 50 mL/sec should be used. As extrapolated Pcrit values including and excluding flow-limited breaths were not clinically different from one another, identification of flow-limited breaths does not appear to be critical to these calculations after underlying criteria for peak inspiratory flow values are met.

DISCLOSURE STATEMENT

The study was supported by NIH grants R01HL085695, R01HL58585, R01HL089447 and P01HL094307. Philips Respironics provided use of the Pcrit device for this study. Dr. Schwab has consulted for ApniCure Foramis Medical Group, and CryOSA. Dr. Pien receives royalties from UpToDate. Dr. Marcus has received research support from Philips Resprionics. The other authors have indicated no financial conflicts of interest. The work for this study was performed at the University of Pennsylvania Perelman School of Medicine, Philadelphia, PA.