TY - JOUR
AU - Wellman,, Andrew
AB - Abstract Study Objectives Oral appliance therapy is an increasingly common option for treating obstructive sleep apnea (OSA) in patients who are intolerant to continuous positive airway pressure (CPAP). Clinically applicable tools to identify patients who could respond to oral appliance therapy are limited. Methods Data from three studies (N = 81) were compiled, which included two sleep study nights, on and off oral appliance treatment. Along with clinical variables, airflow features were computed that included the average drop in airflow during respiratory events (event depth) and flow shape features, which, from previous work, indicates the mechanism of pharyngeal collapse. A model was developed to predict oral appliance treatment response (>50% reduction in apnea–hypopnea index [AHI] from baseline plus a treatment AHI <10 events/h). Model performance was quantified using (1) accuracy and (2) the difference in oral appliance treatment efficacy (percent reduction in AHI) and treatment AHI between predicted responders and nonresponders. Results In addition to age and body mass index (BMI), event depth and expiratory “pinching” (validated to reflect palatal prolapse) were the airflow features selected by the model. Nonresponders had deeper events, “pinched” expiratory flow shape (i.e. associated with palatal collapse), were older, and had a higher BMI. Prediction accuracy was 74% and treatment AHI was lower in predicted responders compared to nonresponders by a clinically meaningful margin (8.0 [5.1 to 11.6] vs. 20.0 [12.2 to 29.5] events/h, p < 0.001). Conclusions A model developed with airflow features calculated from routine polysomnography, combined with age and BMI, identified oral appliance treatment responders from nonresponders. This research represents an important application of phenotyping to identify alternative treatments for personalized OSA management. OSA, oral appliance therapy, upper airway Statement of Significance Treatment response to oral appliance in patients with obstructive sleep apnea can be predicted at baseline from metrics derived from routine polysomnography. Introduction Oral appliance therapy using mandibular advancement device treatment has become an increasingly common option for treating obstructive sleep apnea (OSA) in patients who prefer an oral appliance or are intolerant to continuous positive airway pressure (CPAP) [1, 2]. The advantage of oral appliance therapy is its higher adherence rate compared to CPAP [3–6]. However, oral appliance therapy is limited by its lower efficacy, achieving a complete response in about half of unselected patients with OSA, according to a recent review on this topic [7]. Given the uncertainty of this treatment efficacy, it is difficult for clinicians to select oral appliance therapy over CPAP. One potential solution is the development of clinically applicable tools to identify patients who will achieve a complete response from oral appliance therapy. Clinicians commonly rely on the apnea–hypopnea index (AHI) and demographic data to select patients for oral appliance therapy, despite having little predictive value [8–14]. Physiological research shows that the key to determine whether individuals respond to oral appliance therapy lies in the phenotypic traits of OSA captured from physiological polysomnography (PSG) [15], and site/mechanism of pharyngeal collapse using multisensor catheters [16], awake nasendoscopy [17, 18], drug-induced sleep endoscopy (DISE) [19, 20], and craniofacial measurements using cephalometry [21–24]. In particular, compared to oral appliance nonresponders, responders had a less collapsible pharynx and a more stable respiratory control system (i.e. lower loop gain) [15]. Furthermore, the pharyngeal collapse was more likely to involve the tongue base in oral appliance responders [19, 25], while nonresponders experienced collapse at the palate [19]. These studies were integral in identifying physiological differences between responders and nonresponders to oral appliance treatment and have demonstrated success in predicting oral appliance treatment response [17, 18, 24, 26]. However, they require specialized equipment and expertise, which could impede widespread implementation in the clinical setting. An alternative and potentially more convenient approach would predict oral appliance response from routine PSG. Recently, in our laboratory we have developed noninvasive approaches using signals from routine PSG to detect collapse at the palate [27], tongue base [28], and epiglottis [29], as well as to characterize “collapsibility” [30, 31] (the propensity for airway collapse). These approaches have yet to be applied to predicting oral appliance treatment response. The present study describes a method for predicting oral appliance treatment response using airflow data acquired from routine PSG. In our approach, we assessed (1) several candidate “flow shape” parameters capturing possible sites of airway collapse adapted from methods previously described [27–29, 32] and (2) percent reduction in airflow during respiratory events, i.e. event depth, as a surrogate for collapsibility [30, 33]. Features were limited to those that could be obtained from either in-laboratory or home sleep studies without reliance on electroencephalogram or arousal scoring. Using a predictive modeling approach, we automatically selected the best features to build and validate a simple regression model that predicts response to oral appliance therapy. We also tested the specific hypotheses that favorable responses to oral appliance therapy are associated with less-severe collapsibility, the presence of tongue base collapse, the absence of palatal collapse, and the absence of epiglottic collapse, all measured noninvasively. Methods Individual data from 81 subjects were compiled from three separate studies that included one night off and one night on treatment with the oral appliance: (1) “MADOX” (NCT03189173), (2) “SSPO” [25], and (3) “PROMAD” [19, 34]. PROMAD (n = 36) is recently published [19, 34] and SSPO (n = 25) is pending publication [25]; MADOX data were taken from 20 patients completed at interim analysis. Parent studies were selected based on the availability of unfiltered airflow signals (per American Academy of Sleep Medicine criteria). All studies included patients in the age range between 21 and 70 years and excluded patients with heart failure, central sleep apnea, periodontal disease, insufficient teeth (<8 teeth in each maxillary and mandibular arch), and temporomandibular joint dysfunction. AHI threshold for study inclusion varied across the three studies. Patients with an AHI greater than 20 (MADOX), 10 (SSPO), or 15 (PROMAD) events/h were included per parent study criteria. All patients available from these studies were included in the present analysis. Written informed consent was provided by all patients and approval was granted from their respective Institutional Review Boards. Protocol In two studies (SSPO and MADOX), participants underwent a baseline sleep study and a follow-up sleep study with oral appliance treatment spaced at least 1 week apart [25] (treatment durations 1 and 2 nights, respectively). Participants in these protocols were instructed to sleep supine. In PROMAD, diagnosed patients with OSA underwent a baseline sleep study and a follow-up sleep study on treatment 3 months later (treatment duration 3 months) [19, 34]. Participants in this protocol were not restricted to sleeping supine. Oral appliance All oral appliance devices used were commercially available duobloc titratable mandibular advancement splints. In two protocols (MADOX and SSPO), participants were titrated to their maximum comfortable protrusion with the goal of at least 70% of the maximum protrusion (32 of 45 patients used BluePro, BlueSom, France; 13 of 45 used other custom-made devices). In PROMAD, patients were fitted with a custom-made device (Respident Butterfly, RespiDent, Orthodontic Clinics NV, Antwerp, Belgium) set at 75% of maximum protrusion [19, 34]. Heterogeneity of treatment duration and device type was considered beneficial to ensure the future generalizability of any predictive model across circumstances. Instrumentation On both study nights and in all protocols, standard techniques and criteria were used to score sleep stages and arousals. In two of the protocols, airflow was measured via a sealed nasal mask (SSPO baseline and treatment nights) [25] or oronasal mask (MADOX baseline night only) connected to a pneumotachometer (Hans Rudolph, Kansas City, MO; Validyne, Northridge, CA). In PROMAD baseline and treatment nights [19, 34] and MADOX treatment night, airflow was measured via a nasal pressure cannula. All studies scored apneas as a greater than 90% reduction in airflow and hypopneas as requiring a 3% oxygen desaturation or an arousal from sleep [19, 34]. Hypopneas were based on a 30% reduction in airflow, which was equivalent to the 50% reduction criteria used for nasal pressure. Definition of treatment response Treatment response was defined as a greater than 50% reduction in AHI from baseline plus a treatment AHI less than 10 events/h [11]. Alternative definitions were also assessed in sensitivity analyses (see Supplementary Materials). Noninvasive pathophysiological measures Several studies suggest that a greater drop in airflow during respiratory events (greater “event depth”) reflects greater collapsibility [30, 31, 33], which is expected to reduce oral appliance therapy efficacy [15, 35, 36]. Research from our laboratory also suggests that certain flow shapes reflect the mechanism or site of pharyngeal collapse during sleep. For example, some patients exhibit “pinching” in the expiratory flow shape which reflects a sudden reduction in expiratory airflow due to the prolapse of the palate, causing it to block the nasopharynx and shunt airflow out the mouth [27]. This flow pattern is prevalent in patients with isolated palatal collapse, i.e. palatal collapse without the involvement of the tongue or other structures. Additionally, the presence of a “discontinuity” in inspiratory flow is a characteristic pattern that occurs during the collapse of the epiglottis [29]. Lastly, minimal “scooping” of the inspiratory flow shape is seen in patients with tongue-based obstruction [25, 28] (i.e. scooping reflects a partial collapse of a structure with high dynamic compliance, unlike the larger mass of the tongue). Event depth As a simplified surrogate for collapsibility, event depth was calculated from the mean ventilation profile of the average respiratory event and is described in detail in Figure 1. This method of extracting measures from the average event profile was first described by Azarbarzin et al. [37]. Event depth was calculated as the mean reduction in ventilation from eupnea during the events occurring during sleep. The ventilation signal is strongly correlated between pneumotach flow (R2 = 0.84) and no bias (nasal pressure ventilation/pneumotach ventilation = 1) [32]. Thus, ventilation and measures thereof (i.e. Event Depth) can be reliably computed in datasets with flow measured via pneumotachometer or nasal pressure. Figure 1. Open in new tabDownload slide Representative data from two subjects, one with deep events (A) and another with shallow events (B), to facilitate the explanation of calculating event depth from airflow. The calculation was performed by (i) identifying all scored respiratory events (apneas and hypopneas), (ii) calculating a “ventilation profile” for each event as the uncalibrated volume × respiratory rate, expressed as a percentage of eupneic [mean] ventilation, (iii) aligning the ventilation profiles for each event at event termination (dashed vertical line) and then ensemble averaging. Panel (iv) of A and B illustrates the ensemble-averaged ventilation profile (thick black line) for each subject, superimposed over the ventilation profiles for all respiratory events (thin gray lines). Event depth is calculated as the mean reduction (from eupnea) in ensemble-averaged ventilation during the event. Figure 1. Open in new tabDownload slide Representative data from two subjects, one with deep events (A) and another with shallow events (B), to facilitate the explanation of calculating event depth from airflow. The calculation was performed by (i) identifying all scored respiratory events (apneas and hypopneas), (ii) calculating a “ventilation profile” for each event as the uncalibrated volume × respiratory rate, expressed as a percentage of eupneic [mean] ventilation, (iii) aligning the ventilation profiles for each event at event termination (dashed vertical line) and then ensemble averaging. Panel (iv) of A and B illustrates the ensemble-averaged ventilation profile (thick black line) for each subject, superimposed over the ventilation profiles for all respiratory events (thin gray lines). Event depth is calculated as the mean reduction (from eupnea) in ensemble-averaged ventilation during the event. Flow shape features The primary “pinching,” “discontinuity,” and “scooping” features were calculated as described previously [27–29, 32]. For each subject, the value of each flow shape feature was calculated as the mean value from all breaths that appeared in the margins of a scored hypopnea. All flow shapes included in the present study have been validated to be well characterized by nasal pressure (R2 > 0.5 vs. pneumotach flow) [32]. As with event depth, flow shapes can be reliably computed in datasets with flow measured via pneumotachometer or nasal pressure. Prediction model development Features In addition to event depth and the primary flow shape features, an additional 58 flow shape features (described previously in refs. [27–29, 32]) were made available for selection. In addition, clinical variables (age, sex, body mass index [BMI], neck circumference, and baseline total AHI) were included in a candidate feature set, yielding a total of 67 features. Each candidate variable was pre-transformed (squared, square-root, or cube-root transforms) to provide a normal distribution before analysis. Model development First, to reduce the 67 features to a manageable number, feature selection was performed by conducting bivariate logistic regression models (treatment response vs. each feature), keeping only the top six features since this was considered a sensible number to allow approximately 3 responders and10 nonresponders per feature. Second, from this narrowed pool of six features, forward stepwise logistic regression was performed. Features were kept if they significantly improved the model (p < 0.05, Fisher F-test). Finally, receiver operating characteristic analysis was used to select the optimal cutoff (maximizing sensitivity plus specificity). Model performance Performance was assessed using a leave-one-out cross-validation approach, which is a standard and more conservative method for computing model performance. Briefly, the entire model development process above (including feature selection) was repeated 81 times, i.e. leaving out subject 1, developing a modified model (on 80 subjects), and predicting the outcome of this left-out individual, then repeating for subjects 2–81. We also report the predictive value without cross-validation for the purpose of comparison with past studies (all of which reported accuracy without cross-validation [11]). For performance testing, all models were adjusted for the study site from which they came. Model comparisons In addition to the primary model (above), we also developed two additional models using (1) airflow features alone and (2) clinical features alone, to determine the extent to which airflow phenotype measures predict responses beyond existing clinical features. The models developed were compared using the F-test. Statistical analysis Descriptive characteristics are presented as medians interquartile range (IQR). Distributions of all variables were tested for normality (Shapiro–Wilk test) and transformed if necessary. A comparison of descriptive characteristics and sleep data between responders and nonresponders were tested using the unpaired t-test. Data analysis was performed using MATLAB (Statistics and Machine Learning Toolbox; MathWorks, Natick, MA). Statistical significance was considered at p < 0.05. As the primary test of the specific hypotheses, we performed bivariate logistic regression to model oral appliance response as a function of four separate primary predictors representing collapsibility (event depth), and collapse at the tongue base (negative-effort dependence [28]), epiglottis (inspiratory discontinuity [29]), and palate (expiratory pinching [27]) using a p-threshold of 0.0125 to handle multiple comparisons. Further exploratory analysis (multivariate logistic regression) adjusted for age, sex, BMI, and baseline AHI as potential confounders. To explore a possible confounding effect of the study site (thus also combined effects of treatment duration or airflow recording type), we also adjusted for the site at which data were collected, i.e. Brigham and Women’s Hospital (MMOAT and MADOX) versus University of Antwerp Hospital (PROMAD). As the primary test of prediction model performance, cross-validated accuracy (weighted based on average of sensitivity and specificity to avoid dependence on the prevalence of responders) was compared to the chance value; standard errors and p-values were estimated using the normal approximation to the binomial distribution. Values of sensitivity, specificity, positive predictive value, and negative predictive value were also provided. As a secondary performance test, the difference in oral appliance treatment efficacy (percent reduction in AHI and treatment AHI) between predicted responders and nonresponders (using cross-validated group allocations) was compared (unpaired t-tests). Results Of the 81 patients studied, 31 met the criteria for oral appliance responders. Differences between responders and nonresponders in baseline demographics and AHI are described in Table 1. Notably, nonresponders were significantly older by 4 years, had a higher BMI by 3 kg/m2, a larger neck circumference by 3 cm, and a nonsignificant trend for higher AHI by 8 events/h on baseline PSG. Table 1. Patient demographics and AHI data for all patients and separated into responders and nonresponders (median [IQR]) Variable . All . Nonresponder . Responder . p . Age (years) 50 [45, 56] 53 [46, 61] 49 [41, 53] 0.010 Sex (women/total) 20/81 14/51 6/30 0.453 BMI (kg/m2) 30.3 [26.8, 33.7] 31.7 [27.5, 35.4] 28.6 [25.8, 30.7] 0.006 Neck circumference (cm) 41.0 [38.8, 43.6] 42.1 [39.0, 44.0] 39.1 [38.0, 41.0] 0.017 Study site (Na:Nb) 45:36 37:23 18:13 0.839 Severe OSA (severe/total) 44/81 30/51 14/30 0.289 Total AHI, baseline (/h) 34.0 [22.0, 53.7] 36.3 [23.2, 56.5] 28 [21.2, 41.4] 0.077 Total AHI, treatment (/h) 13.2 [7.9, 25.1] 22.4 [14.1, 29.6] 6.4 [4.0, 8.6] <0.001 % Reduction 54.9 [37.6, 75.9] 43.3 [7.9, 54.2] 84.1 [70.0, 88.4] <0.001 Variable . All . Nonresponder . Responder . p . Age (years) 50 [45, 56] 53 [46, 61] 49 [41, 53] 0.010 Sex (women/total) 20/81 14/51 6/30 0.453 BMI (kg/m2) 30.3 [26.8, 33.7] 31.7 [27.5, 35.4] 28.6 [25.8, 30.7] 0.006 Neck circumference (cm) 41.0 [38.8, 43.6] 42.1 [39.0, 44.0] 39.1 [38.0, 41.0] 0.017 Study site (Na:Nb) 45:36 37:23 18:13 0.839 Severe OSA (severe/total) 44/81 30/51 14/30 0.289 Total AHI, baseline (/h) 34.0 [22.0, 53.7] 36.3 [23.2, 56.5] 28 [21.2, 41.4] 0.077 Total AHI, treatment (/h) 13.2 [7.9, 25.1] 22.4 [14.1, 29.6] 6.4 [4.0, 8.6] <0.001 % Reduction 54.9 [37.6, 75.9] 43.3 [7.9, 54.2] 84.1 [70.0, 88.4] <0.001 BMI, body mass index; Na:Nb, number of subjects from each study site; AHI, apnea–hypopnea index. p-values represent statistical differences between responders and nonresponders. Open in new tab Table 1. Patient demographics and AHI data for all patients and separated into responders and nonresponders (median [IQR]) Variable . All . Nonresponder . Responder . p . Age (years) 50 [45, 56] 53 [46, 61] 49 [41, 53] 0.010 Sex (women/total) 20/81 14/51 6/30 0.453 BMI (kg/m2) 30.3 [26.8, 33.7] 31.7 [27.5, 35.4] 28.6 [25.8, 30.7] 0.006 Neck circumference (cm) 41.0 [38.8, 43.6] 42.1 [39.0, 44.0] 39.1 [38.0, 41.0] 0.017 Study site (Na:Nb) 45:36 37:23 18:13 0.839 Severe OSA (severe/total) 44/81 30/51 14/30 0.289 Total AHI, baseline (/h) 34.0 [22.0, 53.7] 36.3 [23.2, 56.5] 28 [21.2, 41.4] 0.077 Total AHI, treatment (/h) 13.2 [7.9, 25.1] 22.4 [14.1, 29.6] 6.4 [4.0, 8.6] <0.001 % Reduction 54.9 [37.6, 75.9] 43.3 [7.9, 54.2] 84.1 [70.0, 88.4] <0.001 Variable . All . Nonresponder . Responder . p . Age (years) 50 [45, 56] 53 [46, 61] 49 [41, 53] 0.010 Sex (women/total) 20/81 14/51 6/30 0.453 BMI (kg/m2) 30.3 [26.8, 33.7] 31.7 [27.5, 35.4] 28.6 [25.8, 30.7] 0.006 Neck circumference (cm) 41.0 [38.8, 43.6] 42.1 [39.0, 44.0] 39.1 [38.0, 41.0] 0.017 Study site (Na:Nb) 45:36 37:23 18:13 0.839 Severe OSA (severe/total) 44/81 30/51 14/30 0.289 Total AHI, baseline (/h) 34.0 [22.0, 53.7] 36.3 [23.2, 56.5] 28 [21.2, 41.4] 0.077 Total AHI, treatment (/h) 13.2 [7.9, 25.1] 22.4 [14.1, 29.6] 6.4 [4.0, 8.6] <0.001 % Reduction 54.9 [37.6, 75.9] 43.3 [7.9, 54.2] 84.1 [70.0, 88.4] <0.001 BMI, body mass index; Na:Nb, number of subjects from each study site; AHI, apnea–hypopnea index. p-values represent statistical differences between responders and nonresponders. Open in new tab Primary determinants of treatment responses Favorable oral appliance treatment response was associated with Event Depth (i.e. collapsibility) and Expiratory Pinching (i.e. palate-based obstruction), but not with features associated with collapse at the tongue (negative-effort dependence) and epiglottis (inspiratory discontinuity) (Table 2). Adjusting for age, sex, BMI, and baseline AHI did not reduce the strength of the associations for Event Depth and Expiratory Pinching (Table 2). Table 2. Bivariate associations between key traits and response to oral appliance therapy using logistic regression Trait . β [95% CI] . p . Event depth (collapsibility) ‒0.75 [‒1.3 to ‒0.2] 0.006* Expiratory pinching (palate) 0.74 [0.2 to 1.3] 0.005* Negative-effort dependence (tongue base) 0.01 [‒0.3 to 0.3] 0.966 Discontinuity, D1 (epiglottis) ‒0.03 [‒0.1 to 0.1] 0.570 Trait . β [95% CI] . p . Event depth (collapsibility) ‒0.75 [‒1.3 to ‒0.2] 0.006* Expiratory pinching (palate) 0.74 [0.2 to 1.3] 0.005* Negative-effort dependence (tongue base) 0.01 [‒0.3 to 0.3] 0.966 Discontinuity, D1 (epiglottis) ‒0.03 [‒0.1 to 0.1] 0.570 Adjusting for age, sex, and BMI did not reduce the strength of the associations for Event Depth (β: ‒0.73 [‒1.4, ‒0.1], p = 0.024) and Expiratory Pinching (β: 0.70 [0.1, 1.3], p = 0.030). *Bivariate associations for Event Depth and Expiratory Pinching remained significant after Bonferroni correction (p-threshold = 0.0125). Open in new tab Table 2. Bivariate associations between key traits and response to oral appliance therapy using logistic regression Trait . β [95% CI] . p . Event depth (collapsibility) ‒0.75 [‒1.3 to ‒0.2] 0.006* Expiratory pinching (palate) 0.74 [0.2 to 1.3] 0.005* Negative-effort dependence (tongue base) 0.01 [‒0.3 to 0.3] 0.966 Discontinuity, D1 (epiglottis) ‒0.03 [‒0.1 to 0.1] 0.570 Trait . β [95% CI] . p . Event depth (collapsibility) ‒0.75 [‒1.3 to ‒0.2] 0.006* Expiratory pinching (palate) 0.74 [0.2 to 1.3] 0.005* Negative-effort dependence (tongue base) 0.01 [‒0.3 to 0.3] 0.966 Discontinuity, D1 (epiglottis) ‒0.03 [‒0.1 to 0.1] 0.570 Adjusting for age, sex, and BMI did not reduce the strength of the associations for Event Depth (β: ‒0.73 [‒1.4, ‒0.1], p = 0.024) and Expiratory Pinching (β: 0.70 [0.1, 1.3], p = 0.030). *Bivariate associations for Event Depth and Expiratory Pinching remained significant after Bonferroni correction (p-threshold = 0.0125). Open in new tab To explore the robustness of these associations, separate assessments within the two sites (Antwerp and Brigham) were performed. We found that Expiratory Pinching appeared associated with treatment responses at both sites (Antwerp: β = 1.0 [95% confidence interval [CI]: 0.0 to 2.0], p = 0.049; Brigham: β = 0.8 [0.1 to 1.6], p = 0.023). Event Depth was associated with treatment responses at Antwerp (β = −1.6 [‒3.1 to −0.2], p = 0.027) and a trend was observed at Brigham (β = ‒0.6[−1.3 to 0.1], p = 0.11). Prediction model performance Features included in the full model by feature selection were Expiratory Pinching, Event Depth, BMI, and Age (Table 3 and Figure 2). Prediction accuracy of this model via a cross-validation approach was 74 ± 5% (p < 0.001). Compared to predicted nonresponders, predicted responders had significantly higher median percent reduction (64% vs. 48%) and a lower treatment AHI (8 vs. 20 events/h) (Table 4). Without cross-validation (for comparison with past studies [11]) accuracy was 80 ± 5%, with a 24% difference in median percent reduction in AHI between predicted responders and nonresponders (Table 4). Table 3. Parameters for full model, baseline model, and model including only new features Parameters . Standardized β . Odds ratio [95% CI] . p . F-test, p . R2 . Clinical model  Intercept −0.68 0.012 0.002 0.20  Age −0.97 2.64 [1.39 to 5.04] 0.003  BMI −0.83 2.29 [1.30 to 4.05] 0.004 Flow model  Intercept −0.68 0.013 0.002 0.25  Expiratory pinching 0.93 2.53 [1.39 to 4.62] 0.002  Event depth −0.93 2.54 [1.40 to 4.63] 0.002 Full model  Intercept −0.76 0.011 <0.001 0.35  Expiratory pinching 0.85 2.34 [1.19 to 4.59] 0.014  Event depth −0.75 2.13 [1.14 to 3.97] 0.018  BMI −0.73 2.08 [1.06 to 4.07] 0.033  Age −0.79 2.21 [1.17 to 4.19] 0.015 Parameters . Standardized β . Odds ratio [95% CI] . p . F-test, p . R2 . Clinical model  Intercept −0.68 0.012 0.002 0.20  Age −0.97 2.64 [1.39 to 5.04] 0.003  BMI −0.83 2.29 [1.30 to 4.05] 0.004 Flow model  Intercept −0.68 0.013 0.002 0.25  Expiratory pinching 0.93 2.53 [1.39 to 4.62] 0.002  Event depth −0.93 2.54 [1.40 to 4.63] 0.002 Full model  Intercept −0.76 0.011 <0.001 0.35  Expiratory pinching 0.85 2.34 [1.19 to 4.59] 0.014  Event depth −0.75 2.13 [1.14 to 3.97] 0.018  BMI −0.73 2.08 [1.06 to 4.07] 0.033  Age −0.79 2.21 [1.17 to 4.19] 0.015 BMI, body mass index. Standardized β and odds ratios are calculated per SD change. A negative beta coefficient (β) indicates that odds of treatment response increases with a reduction in the input variable (e.g. odds of treatment response increases with lower event depth, age, and BMI). Adjusting each model for study site had no impact on the statistical significance or strength of the associations of the predictor variables with oral appliance treatment response. Open in new tab Table 3. Parameters for full model, baseline model, and model including only new features Parameters . Standardized β . Odds ratio [95% CI] . p . F-test, p . R2 . Clinical model  Intercept −0.68 0.012 0.002 0.20  Age −0.97 2.64 [1.39 to 5.04] 0.003  BMI −0.83 2.29 [1.30 to 4.05] 0.004 Flow model  Intercept −0.68 0.013 0.002 0.25  Expiratory pinching 0.93 2.53 [1.39 to 4.62] 0.002  Event depth −0.93 2.54 [1.40 to 4.63] 0.002 Full model  Intercept −0.76 0.011 <0.001 0.35  Expiratory pinching 0.85 2.34 [1.19 to 4.59] 0.014  Event depth −0.75 2.13 [1.14 to 3.97] 0.018  BMI −0.73 2.08 [1.06 to 4.07] 0.033  Age −0.79 2.21 [1.17 to 4.19] 0.015 Parameters . Standardized β . Odds ratio [95% CI] . p . F-test, p . R2 . Clinical model  Intercept −0.68 0.012 0.002 0.20  Age −0.97 2.64 [1.39 to 5.04] 0.003  BMI −0.83 2.29 [1.30 to 4.05] 0.004 Flow model  Intercept −0.68 0.013 0.002 0.25  Expiratory pinching 0.93 2.53 [1.39 to 4.62] 0.002  Event depth −0.93 2.54 [1.40 to 4.63] 0.002 Full model  Intercept −0.76 0.011 <0.001 0.35  Expiratory pinching 0.85 2.34 [1.19 to 4.59] 0.014  Event depth −0.75 2.13 [1.14 to 3.97] 0.018  BMI −0.73 2.08 [1.06 to 4.07] 0.033  Age −0.79 2.21 [1.17 to 4.19] 0.015 BMI, body mass index. Standardized β and odds ratios are calculated per SD change. A negative beta coefficient (β) indicates that odds of treatment response increases with a reduction in the input variable (e.g. odds of treatment response increases with lower event depth, age, and BMI). Adjusting each model for study site had no impact on the statistical significance or strength of the associations of the predictor variables with oral appliance treatment response. Open in new tab Table 4. Predictive performance of all models with and without cross-validation, as well as median percent reduction in AHI and treatment AHI for predicted responders and nonresponders . Sensitivity . Specificity . PPV . NPV . % Reduction AHI Median [IQR] . . Treatment AHI Median [IQR] (events/h) . . . . . . . Nonresponders . Responders . Nonresponders . Responders . Cross-validation  Clinical model 83% (25/30) 57% (29/51) 53% (25/47) 85% (29/34) 47.5 [25.6, 67.3] 63.0* [42.4, 84.3] 24.0 [12.1, 30.9] 9.5*** [6.1, 17.2]  Flow model 63% (19/30) 78% (40/51) 63% (19/30) 78% (40/51) 45.1 [14.3, 69.5] 67.7** [52.6, 86.8] 16.7 [10.4, 28.9] 8.3** [5.3, 19.1]  Full model 70% (21/30) 78% (40/51) 66% (21/32) 82% (40/49) 48.4 [25.6, 70.2] 64.2** [51.0, 86.1] 19.9 [12.2, 29.5] 8.3*** [5.1, 11.6] Without cross-validation  Clinical model 90% (27/30) 63% (32/51) 53% (27/46) 85% (32/48) 45.0 [12.7, 64.0] 64.0** [45.1, 84.7] 23.7 [13.4, 30.3] 9.1*** [6.1, 14.6]  Flow model 70% (21/30) 78% (40/51) 66% (21/32) 82% (40/49) 45.1 [10.0, 68.5] 67.7** [51.0, 85.7] 19 [10.8, 29.5] 8.3*** [5.7, 15.8]  Full model 77% (23/30) 82% (42/51) 72% (23/32) 86% (42/49) 46.6 [21.7, 67.5] 70.2** [53.3, 86.1] 20.0 [11.5, 29.5] 8.0*** [5.1, 13.0] . Sensitivity . Specificity . PPV . NPV . % Reduction AHI Median [IQR] . . Treatment AHI Median [IQR] (events/h) . . . . . . . Nonresponders . Responders . Nonresponders . Responders . Cross-validation  Clinical model 83% (25/30) 57% (29/51) 53% (25/47) 85% (29/34) 47.5 [25.6, 67.3] 63.0* [42.4, 84.3] 24.0 [12.1, 30.9] 9.5*** [6.1, 17.2]  Flow model 63% (19/30) 78% (40/51) 63% (19/30) 78% (40/51) 45.1 [14.3, 69.5] 67.7** [52.6, 86.8] 16.7 [10.4, 28.9] 8.3** [5.3, 19.1]  Full model 70% (21/30) 78% (40/51) 66% (21/32) 82% (40/49) 48.4 [25.6, 70.2] 64.2** [51.0, 86.1] 19.9 [12.2, 29.5] 8.3*** [5.1, 11.6] Without cross-validation  Clinical model 90% (27/30) 63% (32/51) 53% (27/46) 85% (32/48) 45.0 [12.7, 64.0] 64.0** [45.1, 84.7] 23.7 [13.4, 30.3] 9.1*** [6.1, 14.6]  Flow model 70% (21/30) 78% (40/51) 66% (21/32) 82% (40/49) 45.1 [10.0, 68.5] 67.7** [51.0, 85.7] 19 [10.8, 29.5] 8.3*** [5.7, 15.8]  Full model 77% (23/30) 82% (42/51) 72% (23/32) 86% (42/49) 46.6 [21.7, 67.5] 70.2** [53.3, 86.1] 20.0 [11.5, 29.5] 8.0*** [5.1, 13.0] Sens, sensitivity; spec, specificity; PPV, positive predictive value; NPV, negative predictive value. Performance assessed after adjusting for study site. All performance measures were significantly different from chance performance. Asterisks represent differences in % reduction of AHI and treatment AHI between predicted responders and nonresponders. *p < 0.05, **p < 0.01, ***p < 0.001. Open in new tab Table 4. Predictive performance of all models with and without cross-validation, as well as median percent reduction in AHI and treatment AHI for predicted responders and nonresponders . Sensitivity . Specificity . PPV . NPV . % Reduction AHI Median [IQR] . . Treatment AHI Median [IQR] (events/h) . . . . . . . Nonresponders . Responders . Nonresponders . Responders . Cross-validation  Clinical model 83% (25/30) 57% (29/51) 53% (25/47) 85% (29/34) 47.5 [25.6, 67.3] 63.0* [42.4, 84.3] 24.0 [12.1, 30.9] 9.5*** [6.1, 17.2]  Flow model 63% (19/30) 78% (40/51) 63% (19/30) 78% (40/51) 45.1 [14.3, 69.5] 67.7** [52.6, 86.8] 16.7 [10.4, 28.9] 8.3** [5.3, 19.1]  Full model 70% (21/30) 78% (40/51) 66% (21/32) 82% (40/49) 48.4 [25.6, 70.2] 64.2** [51.0, 86.1] 19.9 [12.2, 29.5] 8.3*** [5.1, 11.6] Without cross-validation  Clinical model 90% (27/30) 63% (32/51) 53% (27/46) 85% (32/48) 45.0 [12.7, 64.0] 64.0** [45.1, 84.7] 23.7 [13.4, 30.3] 9.1*** [6.1, 14.6]  Flow model 70% (21/30) 78% (40/51) 66% (21/32) 82% (40/49) 45.1 [10.0, 68.5] 67.7** [51.0, 85.7] 19 [10.8, 29.5] 8.3*** [5.7, 15.8]  Full model 77% (23/30) 82% (42/51) 72% (23/32) 86% (42/49) 46.6 [21.7, 67.5] 70.2** [53.3, 86.1] 20.0 [11.5, 29.5] 8.0*** [5.1, 13.0] . Sensitivity . Specificity . PPV . NPV . % Reduction AHI Median [IQR] . . Treatment AHI Median [IQR] (events/h) . . . . . . . Nonresponders . Responders . Nonresponders . Responders . Cross-validation  Clinical model 83% (25/30) 57% (29/51) 53% (25/47) 85% (29/34) 47.5 [25.6, 67.3] 63.0* [42.4, 84.3] 24.0 [12.1, 30.9] 9.5*** [6.1, 17.2]  Flow model 63% (19/30) 78% (40/51) 63% (19/30) 78% (40/51) 45.1 [14.3, 69.5] 67.7** [52.6, 86.8] 16.7 [10.4, 28.9] 8.3** [5.3, 19.1]  Full model 70% (21/30) 78% (40/51) 66% (21/32) 82% (40/49) 48.4 [25.6, 70.2] 64.2** [51.0, 86.1] 19.9 [12.2, 29.5] 8.3*** [5.1, 11.6] Without cross-validation  Clinical model 90% (27/30) 63% (32/51) 53% (27/46) 85% (32/48) 45.0 [12.7, 64.0] 64.0** [45.1, 84.7] 23.7 [13.4, 30.3] 9.1*** [6.1, 14.6]  Flow model 70% (21/30) 78% (40/51) 66% (21/32) 82% (40/49) 45.1 [10.0, 68.5] 67.7** [51.0, 85.7] 19 [10.8, 29.5] 8.3*** [5.7, 15.8]  Full model 77% (23/30) 82% (42/51) 72% (23/32) 86% (42/49) 46.6 [21.7, 67.5] 70.2** [53.3, 86.1] 20.0 [11.5, 29.5] 8.0*** [5.1, 13.0] Sens, sensitivity; spec, specificity; PPV, positive predictive value; NPV, negative predictive value. Performance assessed after adjusting for study site. All performance measures were significantly different from chance performance. Asterisks represent differences in % reduction of AHI and treatment AHI between predicted responders and nonresponders. *p < 0.05, **p < 0.01, ***p < 0.001. Open in new tab Figure 2. Open in new tabDownload slide The predictive model illustrated in two dimensions (*adjusted for age and BMI) accurately separated treatment responders from nonresponders. Patients with more pinched expirations and deeper events (bottom right region, shaded red) were less likely to respond to oral appliances (red circles). Shallow events without expiratory pinching (top left region, shaded green) were more likely to respond (green circles). Figure 2. Open in new tabDownload slide The predictive model illustrated in two dimensions (*adjusted for age and BMI) accurately separated treatment responders from nonresponders. Patients with more pinched expirations and deeper events (bottom right region, shaded red) were less likely to respond to oral appliances (red circles). Shallow events without expiratory pinching (top left region, shaded green) were more likely to respond (green circles). The model developed from only clinical features (i.e. clinical model) selected age and BMI as significant predictors (Table 3). Overall cross-validation accuracy of this clinical model was 70 ± 5% (p < 0.001), and predicted responders exhibited a significantly higher median percent reduction in AHI compared to nonresponders (additional 16% reduction) and lower treatment AHI by 14 events/h (Table 4). For the model that only included flow features (i.e. flow model), Event Depth and Expiratory Pinching were selected for inclusion in the model (Table 3). Like the baseline model, cross-validated accuracy was 71 ± 5% (p < 0.001), and the difference between responders and nonresponders in the median percent reduction in AHI was 23% and treatment AHI was 11 events/h (Table 4). With respect to model comparison, in addition to a (slightly) higher prediction accuracy, the full model had a substantially greater R2 compared to both the clinical model (0.20 vs. 0.35; F = 8.2, p < 0.001) and the flow model (0.25 vs. 0.35; F = 5.8, p < 0.01). Discussion The current paper demonstrates that the airflow signal from routine PSG can be used to predict response to oral appliance therapy. We showed for the first time that patients who exhibit palate collapse—as assessed noninvasively by a more “pinched” expiratory flow pattern—improve the least on oral appliance therapy. We also demonstrated that patients with deeper respiratory events, indicating a more collapsible pharynx, also improve the least on oral appliance therapy. Notably, validated noninvasive measures of collapse at the tongue base and epiglottis were not associated with oral appliance response. A prediction model that included key flow features, combined with age and BMI, predicted response with a 74% accuracy (Table 3 and Figure 2; 80% before cross-validation). The model is clinically relevant in that predicted responders achieved a median treatment AHI of 8 events/h compared to 20 events/h in predicted nonresponders. Furthermore, the model outperformed previously published models that used clinical [13, 14] or PSG-derived variables [8–10, 12] and performed comparably to specialized approaches like DISE [19, 20] and awake endoscopy [17, 18] (Figure 4). Figure 3. Open in new tabDownload slide Flow trace from two representative patients (A) with expiratory pinching and (B) without expiratory pinching, which look like a normal expiration. Above the flow trace in A is a profile view of the pharynx illustrating a normal inspiration and an expiration with palatal prolapse whereby the palate flips up to block the nasopharynx causing air to be shunted out of the mouth. This appears in the flow shape as a sudden drop in expiratory flow. On the right are enlarged breaths from A and B with annotations of time durations at or above 90% of peak expiratory flow (t90). Flatness is calculated as the ratio of t90 to total expiratory time (texp). Figure 3. Open in new tabDownload slide Flow trace from two representative patients (A) with expiratory pinching and (B) without expiratory pinching, which look like a normal expiration. Above the flow trace in A is a profile view of the pharynx illustrating a normal inspiration and an expiration with palatal prolapse whereby the palate flips up to block the nasopharynx causing air to be shunted out of the mouth. This appears in the flow shape as a sudden drop in expiratory flow. On the right are enlarged breaths from A and B with annotations of time durations at or above 90% of peak expiratory flow (t90). Flatness is calculated as the ratio of t90 to total expiratory time (texp). Figure 4. Open in new tabDownload slide The model developed in the present study (red circle) outperforms previously published models that used predictors acquired from clinical data and PSG (gray circles), and drug-induced sleep endoscopy (DISE, gray triangles); and performed comparably to past models that used predictors acquired from awake endoscopy (AE, gray triangles). Figure 4. Open in new tabDownload slide The model developed in the present study (red circle) outperforms previously published models that used predictors acquired from clinical data and PSG (gray circles), and drug-induced sleep endoscopy (DISE, gray triangles); and performed comparably to past models that used predictors acquired from awake endoscopy (AE, gray triangles). Palatal prolapse as a predictor The present study represents the first application of flow shape analysis for identifying oral appliance treatment response. Each standard deviation decrease in Expiratory Pinching doubled the odds of responding to oral appliance therapy. Expiratory Pinching was significantly associated with response to oral appliance at both sites (Brigham and Antwerp), demonstrating robustness of the predictor to variability in methodologies across sites (i.e. on-treatment PSG after 1 week [Brigham] vs. 3 months [Antwerp], supine-only sleep at Brigham, and pneumotach flow [Brigham] vs. nasal pressure [Antwerp]). Flow shape analysis is valuable because it is a noninvasive approach to describe the mechanism/site of the pharyngeal collapse. For example, the pinched appearance of the expiratory flow shape has been previously validated to represent expiratory flow limitation caused by bulging or prolapse of the palate during expiration, as visualized with endoscopy during natural sleep [27]. As illustrated in Figure 3, when palatal prolapse occurs on expiration, the palate flips up and blocks the nasopharynx. This causes shunting of air out of the mouth, which appears on the flow trace as a sudden drop in expiratory flow (Figure 3). The relationship between expiratory palatal prolapse and oral appliance treatment response suggests that this collapse mechanism/site is less treatable by oral appliance therapy. Our previous work suggests an association between expiratory palatal prolapse and the site of inspiratory collapse [27]. Palatal prolapse was determined to be a flow signature present in patients with inspiratory “isolated” palatal collapse. This form of collapse, determined by endoscopy during natural sleep, is defined as collapse at the level of the palate without the involvement of the tongue or other airway structures (lateral walls or epiglottis). Furthermore, these patients had an “anteriorly located” tongue base, defined as one in which the vallecula was clearly visible (tongue base not touching epiglottis) during end-expiration. It therefore follows that the patients with a pinched expiratory flow pattern likely experienced isolated palatal collapse on inspiration, which might be harder to treat with an oral appliance. We speculate that mandibular protrusion via oral appliance in such patients will move the tongue anteriorly but may have little effect on the palate. Whereas if the palate is posteriorly positioned secondary to the posterior location of the tongue, anterior movement of the tongue could also move the palate anteriorly and increase the velopharyngeal area. Indeed, Okuno et al. [17] recently imaged the airway via awake nasendoscopy and found that doubling the velopharyngeal cross-sectional area with mandibular protrusion with an oral appliance was a strong predictor of treatment response. In other words, patients in whom mandibular protrusion had minimal effect on the velopharyngeal area were harder to treat with an oral appliance. We suspect that these patients overlap with those experiencing isolated palatal collapse described in the present study. These findings are generally concordant with past research investigating the relationship between the site of collapse and oral appliance treatment efficacy. In a small study of 12 patients, 7 of 8 with collapse at the level of the palate, as determined by a multisensor catheter, failed oral appliance therapy. On the contrary, all four patients with oropharyngeal collapse responded (residual AHI <5) [16]. Furthermore, the most recent DISE study from Op de Beeck et al. [19] found that patients with complete concentric collapse at the level of the palate showed higher odds for worsening of OSA with oral appliance therapy, while collapse at the tongue base was predictive of oral appliance treatment response. Overall, the present study adds to the evidence that patients with isolated palatal collapse are less likely to respond to oral appliance therapy. Event depth as a predictor The average drop in airflow during respiratory events was referred to as the “event depth.” We intuit that deeper respiratory events are associated with an airway that is more collapsible. This intuition is grounded in evidence from previous work by Schwartz et al. [33] who found that the active collapsibility, defined by the critical pressure at which the pharynx collapses while dilator muscles are active (i.e. active Pcrit), was highest (i.e. more collapsible) in patients with a majority of apneas, significantly lower in patients with a majority of hypopneas, and lowest in those with primary snoring. These results suggest that patients with deeper events have a more collapsible airway. Furthermore, work from our laboratory previously showed that average peak flow during non-rapid eye movement sleep is strongly correlated with active collapsibility defined by active Pcrit measurements [30]. Taken together, measures of airflow during sleep can describe pharyngeal compromise and are associated with pharyngeal collapsibility. Of note, each standard deviation reduction in mean Event Depth approximately doubled the odds of responding to oral appliance therapy. Therefore, patients with deeper events are less likely to respond, perhaps because oral appliance therapy has a limited effect size on the airway and may not fully open a highly collapsible airway. This is supported by a past study of 12 patients who underwent physiological PSG showing that collapsibility of the pharynx under passive conditions (i.e. when dilator muscles are relatively inactive) was significantly lower in oral appliance responders [15]. Event Depth was less robust to variability in data collection methodology, compared to Expiratory Pinching, given that it was significantly associated with responses in Antwerp, but only demonstrated a nonsignificant trend at the Brigham and Women’s Hospital. Multivariable model The multivariate model for predicting oral appliance treatment response in the present study included measures of Expiratory Pinching and Event Depth, as well as age and BMI. While the cross-validation accuracy of the full model was moderate, we note that the accuracy (without cross-validation as per past studies [11]) of the model in the present study outperforms models from previous studies that also only used PSG [8–10, 12] and clinical [13, 14] variables (Figure 4). Studies that used PSG variables found that supine-predominant OSA and lower total AHI were associated with improved oral appliance treatment response. Nevertheless, they were less predictive of response than the current study, perhaps because AHI (or AHI-dependent metrics) contains little physiological information and therefore cannot characterize underlying pathophysiological determinants of OSA, such as collapsibility. For example, the AHI classifies apneas and hypopneas equally and does not quantify the depth or duration of the respiratory events [38], both of which may be affected by pharyngeal collapsibility. We propose that performance in the present study is improved due to the physiological specificity of the flow-based predictors for describing collapsibility and site of the collapse, both of which have specific implications for oral appliance treatment response. The full model performed comparably to DISE studies aimed at predicting oral appliance treatment response [19, 20] (Figure 4). The latest such study found that patients with collapse primarily at the palate were less likely to respond to oral appliance therapy, while tongue base collapse predicted oral appliance treatment response [19]. The model also performed comparably to awake nasendoscopy studies [17, 18] (Figure 4), the latest of which found that baseline AHI and expansion of the cross-sectional area of velopharynx produced by the oral appliance predicted treatment response with a sensitivity and specificity of 86% and 81%, respectively [17]. The advantage of endoscopy-based techniques is that the measures utilized (e.g. site of collapse and pharyngeal area) adequately describe the mechanism of pharyngeal collapse. However, endoscopy-based techniques are limited in that they require an additional patient visit, special equipment, and expertise, and in the case of DISE, sedation of the patient in the presence of an anesthesiologist. The present study only requires data from routine PSG to predict treatment response, which is expected to greatly reduce the time and expense compared to endoscopy. The clinical model (age and BMI) performed better than anticipated with a 70% prediction accuracy, despite a low R2 of 0.20. The full model only improved accuracy by 4%, but increased R2 significantly and substantially to 0.35. Therefore, despite the modest improvement in prediction accuracy in the full model versus the clinical model, the importance of airflow metrics in explaining oral appliance treatment response is apparent from the substantial increase in explained variance (i.e. R2). Furthermore, given a modest-to-negligible performance of age and BMI in predicting oral appliance response from past studies [13, 14], it appears somewhat fortuitous that clinical variables performed so well in predicting oral appliance response, while explaining so little variance. Limitations There are several limitations to keep in mind when interpreting the results of this study. First, there is heterogeneity in study designs conducted at different sites. To account for the study site, we included a study site as a covariate in the predictive model (Table 3), which was not significant indicating that treatment response was not significantly affected by variability in the study site (e.g. treatment duration, supine-only sleep, and flow measurement). Furthermore, we demonstrated that the Expiratory Pinching was separately associated with oral appliance treatment response at both sites (Brigham and Women’s Hospital and University of Antwerp Hospital); Event Depth demonstrated a consistent trend at both sites. Second, our study used unfiltered (DC-coupled) airflow signals (best practice per AASM guidelines), which we believe are necessary for flow shape analysis (e.g. optional high-pass “baseline removal” filtering of 0.03 Hz distorts signals); we would not expect our results to apply to filtered airflow signal data. Second, while performance of the model is among the best available (Figure 4), cross-validation sensitivity and specificity were both less than 80% which means that there are potentially other parameters we have not measured that may improve prediction. More data and the development of additional features will be necessary to improve prediction accuracy. Third, we did not confirm the pharyngeal mechanism of collapse with endoscopy in these patients, but rather inferred it from the flow shapes previously validated in our laboratory [27–29, 32]. Fourth, the present study does not link to other clinical outcomes of therapy such as sleep quality, blood pressure, or adherence; and the model was developed retrospectively on data that were prospectively collected for other analyses. As such, while the model represents a major developmental step in predicting oral appliance treatment response from routine PSG, an important next step is validation on prospective data and assessing model performance on predicting relevant clinical outcomes. At first glance, our study appears limited by a lower response rate of 38% compared to other published studies, which average to 64% rate of complete response (criteria: treatment AHI <10 events/h) [7]. However, our study patients had more severe OSA than most of the published clinical trials. Specifically, pretreatment AHI in the present study of 34 events/h (median), with 54% of patients having severe OSA, is higher than most of these published studies. Our study most closely resembles that of Petri et al. [39] who studied patients with a pretreatment AHI of 39.1 events/h, had 56% of patients in the severe category, and reported a treatment response rate of 40%. Thus, we believe that given the patient population studied, our response rate is comparable to previously published data. Another drawback of the present study is that loop gain was not included in the model. Loop gain is an important measure because, together with passive collapsibility (i.e. passive Pcrit) measured in a detailed physiology study, resulted in a model that correctly predicted response to oral appliance treatment in 12 of 13 patients [15]. Loop gain and the other phenotypic traits can be derived from routine PSG [31], but require carefully scored arousals (start and end times), which is not routine clinical practice. To maximize the applicability of the model developed, these traits were not included in the analysis. However, the data analyzed in the present study did have adequately scored arousals, making it possible to calculate the phenotypic traits. Including loop gain and arousal threshold in the final model did not explain the additional variance, nor did they affect prediction accuracy (see Supplementary Table S2). As such, including these phenotypic OSA traits in the feature set would not have changed the presented model. We suspect that loop gain did not improve the model, potentially because it was associated with Event Depth (higher loop gain leads to deeper events) and as such loop gain explains similar variance to Event Depth. Lastly, the present study is limited by the assumption that all variables in the model add linearly. This assumes that Expiratory Pinching and Event Depth have an equal impact across age and BMI. Only linear terms were included because the inclusion of all interactions would have increased the size of the feature set to 4422 variables, which is impractical given the sample size of 81 patients. Instead, we opted to explore interactions between variables of interest post hoc to further expand our understanding of these new airflow features. This analysis revealed significant interactions between Age and Expiratory Pinching, as well as BMI and Expiratory Pinching, both of which significantly improved the model R2 from 0.35 to 0.44 (F = 5.9, p < 0.01) (see Supplementary Materials for additional results). As described in Supplementary Figure S3 and Supplementary Table S2, the interactions suggested that Expiratory Pinching predominantly predicted responses in the younger and leaner population. The effect on oral appliance treatment response disappeared in the heavier and older population (see Supplementary Materials). More data are needed to determine if the model with interaction effects is generalizable to new data. Conclusions The present study is the first to show that patients exhibiting pinched expiratory flow shapes (i.e. isolated palate collapse) are less likely to respond to oral appliance therapy. In addition, using a simple and novel measure of event depth, we showed that patients with a more collapsible pharynx also experienced reduced oral appliance treatment efficacy. Both variables, which can be calculated from routine PSG, when combined with age and BMI form a model capable of predicting oral appliance therapy with 74% accuracy and outperform comparable predictive models previously published [11]. This research represents an important step in the application of phenotyping to identify alternative treatment options and encourage personalized management of OSA. Acknowledgments The authors are grateful to BlueSom for providing oral appliance device support (BluePro) for the parent studies SSPO and MADOX. Funding The research project received funding from the National Institutes of Health (R01HL115459, Wellman PI). BAE is supported by the National Heart Foundation of Australia (101167). MM was supported by the Sao Paulo Research Foundation (FAPESP). SAS, LTM, and AA were supported by the American Heart Association (15SDG25890059, 17POST33410436, 19CDA34660137). BlueSom provided oral appliance devices (BluePro) for our studies (SSPO, MADOX) at no cost. Conflict of interest statement. Financial disclosures: OMV reports grants from Government of Flanders (Belgium)-IWT, during the conduct of the study; grants at Antwerp University Hospital from Philips and Somnomed; research support from Inspire Medical Systems; member of Advisory Board for Zephyr and Liva Nova, speaker’s fees from Somnomed and Inspire Medical Systems; OMV holds a Senior Clinical Fellowship Grant (Fundamenteel Klinisch Mandaat) from Research Foundation—Flanders—Vlaanderen (FWO). LTM works as a consultant for Apnimed and received personal fees as a consultant for Cambridge Sound Management outside the submitted work. LTM and AW have a financial interest in Apnimed, a company developing pharmacologic therapies for sleep apnea. Their interests were reviewed and are managed by Brigham and Women’s Hospital and Partners HealthCare in accordance with their conflict of interest policies. AW works as a consultant for Apnimed, Somnifix, and Nox and he has received grants from Somnifix and Sanofi. DPW receives personal fees as a consultant from Apnimed outside the submitted work, and he is the Chief Scientific Officer for Philips Respironics. SAS receives personal fees as a consultant for Cambridge Sound Management, Nox Medical, and Merck outside the submitted work and receives grant support from Apnimed and Prosomnus. AA receives personal fees as a consultant for Somnifix and Apnimed and receives grant support from the American Heart Association, American Academy of Sleep Medicine, and Somnifix. BAE receives grant support from Apnimed and NHMRC. GSH has received equipment for research from Resmed, Philips Respironics, and Air Liquide Healthcare. JV receives personal fees as a consultant or speaker from Ectosense, Philips, ResMed Narval, Sanofi, AstraZen, Daiichi-Sankyo, Zambon, Bioprojet, Schwabe Pharma, Metagenics, Agfa-Gevaert, SomnoMed, Springer, Inspire Medical Systems, Boehringer Ingelheim, Plastiflex, Jansen-Cilag, Vivisol, and Air Liquide outside the submitted work. MB reports a governmental grant from the Government of Flanders (Belgium)-IWT during the conduct of the study. MB is a member of the advisory board of Somnomed and ResMed and receives as such travel funding and speaker fees outside the submitted work. DV, MM, SODB, RR, SAJ, SWK have no conflict of interest to disclose. Nonfinancial disclosures: None. Author Contributions DV contributed to study design, data analysis and interpretation, and drafting and review of the manuscript for important intellectual content. AA contributed to the study design, data analysis and interpretation, and review of the manuscript. MM, SODB, OMV, BAE, RR, GSH, SAJ, JV, and MB contributed to data collection and review of the manuscript. LTM, SWK, and DPW contributed to data interpretation and review of the manuscript. AW contributed to the study design, data analysis and interpretation, and drafting and review of the manuscript. SAS contributed to the study design, data collection, data analysis and interpretation, and drafting and review of the manuscript. References 1. Marklund M , et al. Non-CPAP therapies in obstructive sleep apnoea: mandibular advancement device therapy . Eur Respir J. 2012 ; 39 ( 5 ): 1241 – 1247 . Google Scholar Crossref Search ADS PubMed WorldCat 2. Ramar K , et al. Clinical practice guideline for the treatment of obstructive sleep apnea and snoring with oral appliance therapy: an update for 2015 . J Clin Sleep Med. 2015 ; 11 ( 7 ): 773 – 827 . Google Scholar Crossref Search ADS PubMed WorldCat 3. Phillips CL , et al. Health outcomes of continuous positive airway pressure versus oral appliance treatment for obstructive sleep apnea: a randomized controlled trial . Am J Respir Crit Care Med. 2013 ; 187 ( 8 ): 879 – 887 . Google Scholar Crossref Search ADS PubMed WorldCat 4. Sutherland K , et al. Efficacy versus effectiveness in the treatment of obstructive sleep apnea: CPAP and oral appliances . J Dent Sleep Med . 2015 ; 02 ( 04 ): 175 – 181 . doi:10.15331/jdsm.5120 Google Scholar Crossref Search ADS WorldCat 5. Dieltjens M , et al. Objectively measured vs self-reported compliance during oral appliance therapy for sleep-disordered breathing . Chest. 2013 ; 144 ( 5 ): 1495 – 1502 . Google Scholar Crossref Search ADS PubMed WorldCat 6. Vanderveken OM , et al. Objective measurement of compliance during oral appliance therapy for sleep-disordered breathing . Thorax. 2013 ; 68 ( 1 ): 91 – 96 . Google Scholar Crossref Search ADS PubMed WorldCat 7. Sutherland K , et al. Oral appliance treatment for obstructive sleep apnea: an update . J Clin Sleep Med. 2014 ; 10 ( 2 ): 215 – 227 . Google Scholar Crossref Search ADS PubMed WorldCat 8. Chung JW , et al. Treatment outcomes of mandibular advancement devices in positional and nonpositional OSA patients . Oral Surg Oral Med Oral Pathol Oral Radiol Endod. 2010 ; 109 ( 5 ): 724 – 731 . Google Scholar Crossref Search ADS PubMed WorldCat 9. Holley AB , et al. Efficacy of an adjustable oral appliance and comparison with continuous positive airway pressure for the treatment of obstructive sleep apnea syndrome . Chest. 2011 ; 140 ( 6 ): 1511 – 1516 . Google Scholar Crossref Search ADS PubMed WorldCat 10. Lee CH , et al. The effect of positional dependency on outcomes of treatment with a mandibular advancement device . Arch Otolaryngol Head Neck Surg. 2012 ; 138 ( 5 ): 479 – 483 . Google Scholar Crossref Search ADS PubMed WorldCat 11. Okuno K , et al. Prediction of oral appliance treatment outcomes in obstructive sleep apnea: a systematic review . Sleep Med Rev. 2016 ; 30 : 25 – 33 . Google Scholar Crossref Search ADS PubMed WorldCat 12. Yoshida K . Influence of sleep posture on response to oral appliance therapy for sleep apnea syndrome . Sleep. 2001 ; 24 ( 5 ): 538 – 544 . Google Scholar Crossref Search ADS PubMed WorldCat 13. Tsuiki S , et al. Oropharyngeal crowding and obesity as predictors of oral appliance treatment response to moderate obstructive sleep apnea . Chest. 2013 ; 144 ( 2 ): 558 – 563 . Google Scholar Crossref Search ADS PubMed WorldCat 14. Sutherland K , et al. Oral appliance treatment response and polysomnographic phenotypes of obstructive sleep apnea . J Clin Sleep Med. 2015 ; 11 ( 8 ): 861 – 868 . Google Scholar Crossref Search ADS PubMed WorldCat 15. Edwards BA , et al. Upper-airway collapsibility and loop gain predict the response to oral appliance therapy in patients with obstructive sleep apnea . Am J Respir Crit Care Med. 2016 ; 194 ( 11 ): 1413 – 1422 . Google Scholar Crossref Search ADS PubMed WorldCat 16. Ng AT , et al. Oropharyngeal collapse predicts treatment response with oral appliance therapy in obstructive sleep apnea . Sleep. 2006 ; 29 ( 5 ): 666 – 671 . Google Scholar PubMed OpenURL Placeholder Text WorldCat 17. Okuno K , et al. Endoscopy evaluation to predict oral appliance outcomes in obstructive sleep apnoea . Eur Respir J. 2016 ; 47 ( 5 ): 1410 – 1419 . Google Scholar Crossref Search ADS PubMed WorldCat 18. Chan AS , et al. The effect of mandibular advancement on upper airway structure in obstructive sleep apnoea . Thorax. 2010 ; 65 ( 8 ): 726 – 732 . Google Scholar Crossref Search ADS PubMed WorldCat 19. Op de Beeck S , et al. Phenotypic labelling using drug-induced sleep endoscopy improves patient selection for mandibular advancement device outcome: a prospective study . J Clin Sleep Med . 2019 ; 15 ( 08 ): 1089 – 1099 . Google Scholar Crossref Search ADS PubMed WorldCat 20. Vroegop AV , et al. Sleep endoscopy with simulation bite for prediction of oral appliance treatment outcome . J Sleep Res. 2013 ; 22 ( 3 ): 348 – 355 . Google Scholar Crossref Search ADS PubMed WorldCat 21. Liu Y , et al. Supine cephalometric analyses of an adjustable oral appliance used in the treatment of obstructive sleep apnea . Sleep Breath. 2000 ; 4 ( 2 ): 59 – 66 . Google Scholar Crossref Search ADS PubMed WorldCat 22. Liu Y , et al. Cephalometric and physiologic predictors of the efficacy of an adjustable oral appliance for treating obstructive sleep apnea . Am J Orthod Dentofacial Orthop. 2001 ; 120 ( 6 ): 639 – 647 . Google Scholar Crossref Search ADS PubMed WorldCat 23. Ng AT , et al. Cephalometry and prediction of oral appliance treatment outcome . Sleep Breath. 2012 ; 16 ( 1 ): 47 – 58 . Google Scholar Crossref Search ADS PubMed WorldCat 24. Shen HL , et al. Craniofacial morphologic predictors of oral appliance outcomes in patients with obstructive sleep apnea . J Am Dent Assoc. 2012 ; 143 ( 11 ): 1209 – 1217 . Google Scholar Crossref Search ADS PubMed WorldCat 25. Marques M , et al. Structure and severity of pharyngeal obstruction determine oral appliance efficacy in sleep apnoea . J Physiol. 2019 ; 597 ( 22 ): 5399 – 5410 . Google Scholar Crossref Search ADS PubMed WorldCat 26. Remmers J , et al. Remotely controlled mandibular protrusion during sleep predicts therapeutic success with oral appliances in patients with obstructive sleep apnea . Sleep. 2013 ; 36 ( 10 ): 1517 – 1525, 1525A . Google Scholar Crossref Search ADS PubMed WorldCat 27. Azarbarzin A , et al. Palatal prolapse as a signature of expiratory flow limitation and inspiratory palatal collapse in patients with obstructive sleep apnoea . Eur Respir J . 2018 ; 51 ( 2 ): 1701419 . Google Scholar Crossref Search ADS PubMed WorldCat 28. Genta PR , et al. Airflow shape is associated with the pharyngeal structure causing OSA . Chest. 2017 ; 152 ( 3 ): 537 – 546 . Google Scholar Crossref Search ADS PubMed WorldCat 29. Azarbarzin A , et al. Predicting epiglottic collapse in patients with obstructive sleep apnoea . Eur Respir J . 2017 ; 50 ( 3 ): 1700345 . Google Scholar Crossref Search ADS PubMed WorldCat 30. Azarbarzin A , et al. Estimation of pharyngeal collapsibility during sleep by peak inspiratory airflow . Sleep . 2017 ; 40 ( 1 ). doi:10.1093/sleep/zsw005 OpenURL Placeholder Text WorldCat 31. Sands SA , et al. Phenotyping pharyngeal pathophysiology using polysomnography in patients with obstructive sleep apnea . Am J Respir Crit Care Med. 2018 ; 197 ( 9 ): 1187 – 1197 . Google Scholar Crossref Search ADS PubMed WorldCat 32. Mann DL , et al. Quantifying the magnitude of pharyngeal obstruction during sleep using airflow shape . Eur Respir J . 2019 ; (epub ahead of print:1802262). doi:10.1183/13993003.02262-2018 OpenURL Placeholder Text WorldCat 33. Gleadhill IC , et al. Upper airway collapsibility in snorers and in patients with obstructive hypopnea and apnea . Am Rev Respir Dis. 1991 ; 143 ( 6 ): 1300 – 1303 . Google Scholar Crossref Search ADS PubMed WorldCat 34. Verbruggen AER , et al. Predicting therapeutic outcome of mandibular advancement device treatment in obstructive sleep apnoea (PROMAD): study design and baseline characteristics . J Dent Sleep Med . 2016 ; 03 ( 04 ): 119 – 138 . doi:10.15331/jdsm.6250 Google Scholar Crossref Search ADS WorldCat 35. Tsuiki S , et al. Optimal positive airway pressure predicts oral appliance response to sleep apnoea . Eur Respir J. 2010 ; 35 ( 5 ): 1098 – 1105 . Google Scholar Crossref Search ADS PubMed WorldCat 36. Marklund M , et al. Mandibular advancement devices in 630 men and women with obstructive sleep apnea and snoring: tolerability and predictors of treatment success . Chest. 2004 ; 125 ( 4 ): 1270 – 1278 . Google Scholar Crossref Search ADS PubMed WorldCat 37. Azarbarzin A , et al. The hypoxic burden of sleep apnoea predicts cardiovascular disease-related mortality: the osteoporotic fractures in men study and the sleep heart health study . Eur Heart J. 2019 ; 40 ( 14 ): 1149 – 1157 . Google Scholar Crossref Search ADS PubMed WorldCat 38. Punjabi NM . COUNTERPOINT: is the apnea-hypopnea index the best way to quantify the severity of sleep-disordered breathing? No . Chest. 2016 ; 149 ( 1 ): 16 – 19 . Google Scholar Crossref Search ADS PubMed WorldCat 39. Petri N , et al. Mandibular advancement appliance for obstructive sleep apnoea: results of a randomised placebo controlled trial using parallel group design . J Sleep Res. 2008 ; 17 ( 2 ): 221 – 229 . Google Scholar Crossref Search ADS PubMed WorldCat Author notes Equal contributions. © Sleep Research Society 2020. Published by Oxford University Press on behalf of the Sleep Research Society. All rights reserved. For permissions, please e-mail journals.permissions@oup.com. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)
TI - Predicting sleep apnea responses to oral appliance therapy using polysomnographic airflow
JF - SLEEP
DO - 10.1093/sleep/zsaa004
DA - 2020-07-13
UR - https://www.deepdyve.com/lp/oxford-university-press/predicting-sleep-apnea-responses-to-oral-appliance-therapy-using-JEYzfXe3gi
DP - DeepDyve
ER -