Abstract

Background Differences in the pharyngeal site of collapse influence efficacy of non-continuous positive airway pressure therapies for obstructive sleep apnoea (OSA). Notably, complete concentric collapse at the level of the palate (CCCp) during drug-induced sleep endoscopy (DISE) is associated with reduced efficacy of hypoglossal nerve stimulation, but CCCp is currently not recognisable using polysomnography. Here we develop a means to estimate DISE-based site of collapse using overnight polysomnography.

Methods 182 OSA patients provided DISE and polysomnography data. Six polysomnographic flow shape characteristics (mean during hypopnoeas) were identified as candidate predictors of CCCp (primary outcome variable, n=44/182), including inspiratory skewness and inspiratory scoopiness. Multivariable logistic regression combined the six characteristics to predict clear presence (n=22) versus absence (n=128) of CCCp (partial collapse and concurrent tongue base collapse excluded). Odds ratios for actual CCCp between predicted subgroups were quantified after cross-validation. Secondary analyses examined complete lateral wall, tongue base or epiglottis collapse. External validation was performed on a separate dataset (ntotal=466).

Results CCCp was characterised by greater scoopiness (β=1.5±0.6 per 2sd, multivariable estimate±se) and skewness (β=11.4±2.4) compared with non-CCCp. The odds ratio for CCCp in predicted positive versus negative subgroups was 5.0 (95% CI 1.9–13.1). The same characteristics provided significant cross-validated prediction of lateral wall (OR 6.3, 95% CI 2.4–16.5), tongue base (OR 3.2, 95% CI 1.4–7.3) and epiglottis (OR 4.4, 95% CI 1.5–12.4) collapse. CCCp and lateral wall collapse shared similar characteristics (skewed, scoopy), diametrically opposed to tongue base and epiglottis collapse characteristics. External validation confirmed model prediction.

Conclusions The current study provides a means to recognise patients with likely CCCp or other DISE-based site of collapse categories using routine polysomnography. Since site of collapse influences therapeutic responses, polysomnographic airflow shape analysis could facilitate precision site-specific OSA interventions.

Introduction

Non-continuous positive airway pressure (non-CPAP) treatments for obstructive sleep apnoea (OSA) patients, including hypoglossal nerve stimulation and oral appliance treatment, are characterised by high adherence, yet their efficacy is patient-dependent [1]. An important determinant of suitability of these therapies is the site, pattern and degree of pharyngeal collapse. Most notably, patients with complete concentric collapse at the level of the palate (CCCp) and/or oropharyngeal lateral wall collapse respond less favourably to hypoglossal nerve stimulation or oral appliance therapy than those with tongue base obstruction [2–6], with CCCp being a formal exclusion parameter for hypoglossal nerve stimulation treatment. In clinical practice, this information can be garnered from an advanced procedure known as drug-induced sleep endoscopy (DISE) where pharyngeal collapse characteristics are assessed using a flexible nasopharyngoscope during sedation designed to mimic natural sleep [7, 8]. However, currently the sites of pharyngeal collapse are not discernible via routine polysomnography.

Accumulating evidence supports the notion that different sites of collapse produce recognisably different airflow shape profiles during flow limitation [9–12]. During natural sleep, Genta et al. [11] showed that inspiratory scoopiness, or negative effort dependence (NED), is associated with the site of collapse. Breaths during tongue base collapse showed little scoopiness, palatal or lateral wall collapse had moderate scoopiness, while epiglottis collapse showed high scoopiness [11, 12]. Furthermore, epiglottis collapse has been associated with high jaggedness and an elevated discontinuity index [9], while palatal prolapse during natural sleep was associated with increased expiratory flow limitation [10]. To date, however, no large study has utilised polysomnography to identify the site of collapse measured separately via DISE.

The current study sought to evaluate whether airflow shapes observed in a routine polysomnographic study can provide insight into the likely pharyngeal structures contributing to collapse as seen during DISE. Specifically, we developed a predictive model to identify the site and pattern of collapse using a small number of airflow shape characteristics (supplementary figure E1). Given the clinical implications of having CCCp during DISE, we focused on recognising this specific DISE pattern. We also specifically tested the hypothesis that CCCp during DISE is associated with greater scoopiness on polysomnography.

MethodsParticipants

The current study describes the development of a prediction model using a new retrospective cohort of 182 subjects (“DISEpsg”) from Antwerp University Hospital (Edegem, Belgium) whose clinical data were collected specifically for the current protocol. All 216 candidate adult patients who had DISE between 1 January 2018 and 12 February 2020 and moderate-to-severe OSA (apnoea–hypopnoea index (AHI) >15 events·h−1) on baseline polysomnography within 2 years of the DISE date were identified; 174 had already provided consent that covered the current analysis and an additional eight individuals signed additional informed consent, providing a total of 182 participants who provided written informed consent for analysis. The current study was registered at ClinicalTrials.gov with identifier number NCT04753684. Patients underwent DISE as part of their standard clinical care. DISE was performed as indication for mandibular advancement device, hypoglossal nerve stimulation or other surgical OSA treatment.

Polysomnography protocol

Patients underwent in-lab polysomnography at Antwerp University Hospital as part of routine clinical care (BrainLab RT; Natus Group (OSG), Kontich, Belgium). Patients were fitted with the standard polysomnography equipment including electroencephalography (six leads: F4, C3, C4, O1, M1 and M2), electrooculography, nasal pressure airflow (raw unfiltered), oximetry, breathing effort using respiratory induction plethysmography, body position and muscle activity. Pseudonymised data were exported to EDF file format and analysed using MATLAB 2018a (MathWorks, Natick, MA, USA). Hypopnoeas were scored according to the American Academy of Sleep Medicine 2012 guidelines [13, 14].

DISE protocol

DISE was performed by an experienced ear, nose and throat (ENT) surgeon, independent from the polysomnography measurement. Midazolam sedation (1.5 mg bolus) was maintained using target-controlled propofol infusion (2.0–3.0 µg). A flexible fibreoptic nasopharyngoscope (Olympus END-GP, 3.7 mm diameter; Olympus Europe, Hamburg, Germany) facilitated pharyngeal visualisation. The sites (palate, oropharynx, tongue base, hypopharynx, epiglottis), pattern (anteroposterior, concentric, laterolateral) and degree (partial, complete) of collapse were documented using a standardised scoring system [15]. To minimise inter-rater variability, data were scored prior to (and independently of) polysomnographic flow shape analyses by one ENT surgeon experienced in DISE (E. Van de Perck). For the current analysis, oropharyngeal and hypopharyngeal collapse categories were pooled to describe lateral wall collapse.

Our primary focus was to develop a prediction model with distinct flow shape characteristics that differentiates patients with CCCp (n=44) versus without CCCp. Accordingly: 1) patients with partial concentric collapse were excluded from the primary analysis (n=10 for concentric palate) to reduce ambiguity relating to the presence of CCCp and 2) patients with concomitant CCCp and (complete or partial) tongue base collapse were removed (n=22), since a major goal was to discriminate between CCCp and tongue base collapse thereby initially avoiding complexities relating to multi-level collapse. Notably, we did not exclude lateral wall collapse from CCCp on the basis that they commonly occur together and CCCp may have overlapping lateral wall involvement. The impact of these decisions was carefully examined (see “Assessment of partial collapse” and “Assessment of multi-level collapse” in “Sensitivity analysis – alternative patient and sleep stage criteria” in the supplementary methods and results).

The secondary analyses examined complete lateral wall, tongue base and epiglottis collapse. For each analysis we also sought to minimise ambiguity; partial collapse categories were excluded. Where possible, multi-level collapse was minimised, i.e. for lateral wall collapse we excluded partial and complete tongue base collapse, for tongue base and epiglottis collapse we excluded complete lateral wall collapse and CCCp. See the Results for a summary of the individuals included in each analysis.

Data analysis overview

Our research strategy first involved visually reviewing flow shape characteristics during respiratory events of patients with CCCp versus without CCCp. Incorporating the knowledge gleaned from visual inspection, we manually selected a small number of interpretable flow shape characteristics and used these to develop a regression-based prediction model that identified patients at greatest odds of exhibiting CCCp as observed during DISE.

Flow shape characteristics

Flow shape calculation and model prediction are fully automated and rely minimally on human input. Feature selection that was part of model development was based on a combination of manual (visual) selection, expert opinion and statistical methods. Flow shape characteristics [16] describing aspects of non-rounded airflow behaviour (flattening, scooping, fluttering, within-breath timing features) were automatically calculated, six of which were selected in two phases.

In the first phase, we selected 25 characteristics using a bivariate screening procedure (see “Flow shape characteristics” in the supplementary methods for the complete list). In the second phase, using manual selection, six individual characteristics (table 1) were judiciously selected by investigators based on multiple approaches. First, high-frequency characteristics (snoring, flutter) were excluded because filtering can attenuate snoring in clinical practice. Second, visual inspection of the raw signals of the characteristic hypopnoea events (figure 1; see “Visualisation of characteristic events” in the supplementary methods) was used to visually compare patients with CCCp versus without CCCp; flow shape characteristics had to be interpretable and recognisably different in visual analysis. Likewise, characteristics relating to “scoopiness” (i.e. NED) were prioritised based on prior expectation [11]. Conceptually similar characteristics were avoided (e.g. flatness at different thresholds, multiple definitions of scoopiness/NED). Note, while flow shape characteristics of all individual breaths during hypopnoea were used for model development and during all other steps of feature selection, average flow shapes were constructed for visual inspection. Bivariate analysis results (CCCp versus non-CCCp) were also considered. Six characteristics were considered optimal and manageable from the perspective of future translational application and troubleshooting (see “Sensitivity analysis – number of flow shape characteristics used” in the supplementary methods). The same six characteristics were maintained for the prediction of other sites.

TABLE 1

Shortlist of features that were included in the final analysis

FIGURE 1

Visual inspection of characteristic event (ensemble average flow shape) for a) a representative patient with complete concentric collapse at the level of the palate (CCCp) and b) a patient with tongue base collapse. Patients with CCCp (a) exhibited increased scoopiness (higher negative effort dependence (NED)), increased inspiratory skewness (left-leaning inspiration) and greater early inspiratory peak flow, as shown. Other parameters included in the analyses were rise time, early inspiratory versus expiratory peak time and early inspiratory volume. Ensemble average flow shapes were constructed by ensemble averaging 15 breaths, centred around the scored end of the hypopnoea. Start and end points of each inspiration and expiration of each last hypopnoea breath were determined. Next, inspiration and expiration were averaged and joined back together. This process was repeated for all 15 positions of the ensemble average flow shape. A detailed overview on this technique is presented in “Visualisation of characteristic events” in the supplementary methods.

For each patient, a table was constructed with one row per breath (breaths outside hypopnoeas were excluded) that described the value for each of the six flow shape characteristics; mean values for each characteristic provided a representative value for each patient for analysis (supplementary figure E1).

Statistical analysis

Statistical analysis was performed using MATLAB. Statistical significance was considered if p<0.05.

Primary analysis: CCCp

Multivariable logistic regression modelling combined the six characteristics (continuous independent predictor variables) to predict site of collapse (binary dependent variable). Odds ratios for true CCCp between predicted subgroups (predicted absence and presence of CCCp) were quantified before and after leave-one-patient-out cross-validation to assess model predictive value.

Lateral wall, tongue base and epiglottis collapse

The primary analysis was repeated for the other collapse types, and for the pooled model differentiating patients with CCCp and/or complete lateral wall collapse from patients with tongue base and/or epiglottis collapse.

Night-to-night repeatability

To assess night-to-night repeatability, model-predicted CCCp probability was examined in a separate study, in which 18 patients underwent a baseline and placebo drug study [17]; here, pneumotach airflow was available rather than nasal pressure. Intra-class correlation (absolute agreement method) was used to describe repeatability of individual flow shape characteristics and the resultant model-predicted site of collapse probabilities across the 2 nights.

External validation

For external validation, we calculated model-predicted CCCp probability (plus other sites) in a separate study (“DISEflow”, n=466) performed at Mass Eye and Ear (Boston, MA, USA), in which patients with diagnosed OSA (any severity) underwent simultaneous DISE with pneumotachograph airflow. Logistic regression evaluated the association between true CCCp (binary dependent variable) with model-predicted CCCp probability (independent variable); analyses were repeated for other sites (see “External validation” in the supplementary methods for more details).

Results

Data from all 182 patients were analysed (table 2). Per patient, 148±71 hypopnoea events were included in the analysis, with a total of 706±433 breaths recorded during these hypopnoea events. Regarding DISE subgroups, 44 patients exhibited CCCp, 54 exhibited complete lateral wall collapse, 58 exhibited complete tongue base collapse and 28 exhibited complete epiglottis collapse. After exclusion of partial collapse and concomitant tongue base collapse (for CCCp and lateral wall collapse prediction) the main analytic samples were: CCCp (n=22/150), lateral wall collapse (n=26/104), tongue base collapse (n=37/113) and epiglottis collapse (n=18/158). See table 3 for details.

TABLE 2

Baseline clinical characteristics for all patients (n=182) and the patient subgroups based on collapse type

TABLE 3

Overview of the subjects included for each analysis

Complete prediction modelsPrimary analysis: CCCp

In multivariable regression, CCCp was associated with all six characteristics, notably greater scoopiness (β=6.92±1.44 per 2sd; p<0.0001, as hypothesised), greater skewness (β=11.40±2.43 per 2sd; p<0.0001, i.e. positively skewed/left-leaning), greater early inspiratory peak flow (β=2.20±0.64 per 2sd; p=0.0006), greater rise time (β=1.48±0.75 per 2sd; p=0.047), lower early inspiratory volume (β= −14.44±3.23 per 2sd; p<0.0001) and lower inspiratory versus expiratory peak time ratio (β= −1.50±0.56 per 2sd; p=0.0076) compared with patients without CCCp (pseudo-R2=0.31, model p<0.0001, cross-validated accuracy 0.73±0.04) (table 4; see also supplementary results). Additional adjustment for covariates (AHI, body mass index (BMI) and sex) had minimal effect on these multivariable associations (table 4, CCCp column, likelihood ratio test: p=1.7×10−7versus covariates alone; see supplementary table E3 for adjusted p-values). The odds ratio for CCCp in predicted CCCp versus predicted non-CCCp was 5.0 (95% CI 1.9–13.1) after cross-validation (table 4). Example breaths highlighting how CCCp may be recognised from the flow signal are shown in figure 2. Figure 3a shows how key characteristics contribute to predicted CCCp versus non-CCCp.

TABLE 4

Model parameters, odds ratios and performance after cross-validation for each model.

FIGURE 2

Raw individual breath data of a) patients with complete concentric collapse at the level of the palate (CCCp) and b) patients with tongue base collapse. Overall probabilities of CCCp or tongue base collapse per patient are depicted on the left-hand side of each panel. Probabilities of CCCp or tongue base collapse for individual breaths are depicted below each breath.

FIGURE 3

Simplified two-trait model “slices” for each of the five models: a) complete concentric collapse at the level of the palate (CCCp), b) lateral wall collapse (LW), c) tongue base collapse (TB), d) epiglottis collapse (EG) and e) CCCp/LW versus TB/EG. Each slice plot considers two of six features. The other four features in the model are at their mean value. Slices representing the most important model contributors (inspiratory skewness, rise time, early inspiratory volume and negative effort dependence (NED)) are depicted here. Full model representations are shown in the supplementary material. a–d) Patients with a certain collapse type are depicted in red, patients without this collapse type are depicted in green. Open circles denote patients for whom the model simplification does not apply as the other traits are too far from the mean value for this specific two-dimensional model “slice”. Background colours represent predicted presence of collapse. e) Green dots represent patients with TB or EG, red dots represent patients with CCCp or LW. Background colours depict the predicted site of collapse based on the model (green: TB or EG; red: CCCp or LW).

Secondary analyses: lateral wall, tongue base and epiglottis collapse

Separate models using the same flow shape characteristics provided promising prediction of complete lateral wall (n=26/104; cross-validated OR 6.3, 95% CI 2.4–16.5), tongue base (n=37/113; OR 3.2, 95% CI 1.4–7.3) or epiglottis (n=18/158; OR 4.4, 95% CI 1.5–12.4) collapse. Characteristics between CCCp and lateral wall collapse were similar (scoopy, left-leaning), and diametrically opposed to tongue base and epiglottis characteristics (table 4 and figure 3b–d).

An exploratory model discriminated between CCCp or lateral wall versus tongue base or epiglottis collapse (n=33/34; OR 7.5, 95% CI 2.5–22.1; pseudo-R2=0.46) (table 4 and figure 3e).

Figures including model slices and detailed tables with unstandardised β-values and probability cut-offs for all six characteristics of all models are included in the supplementary material (supplementary tables E3–E12 and supplementary figures E2–E6).

Night-to-night repeatability analysis

In a separate sample (n=18, 2 nights ∼1–4 weeks apart, median (interquartile range) AHI 52 (24–75) events·h−1, BMI 31 (27–35) kg·m−2, age 45 (35–51) years [17]), intra-class correlations (ICCs) were 0.75–0.90 for the six individual flow shape characteristics (figure 4a) and 0.71–0.81 for site of collapse probability scores (ICC 0.76 for CCCp) (figure 4b) (see supplementary table E14).

FIGURE 4

Repeatability analysis on a different cohort (n=18 subjects, n=36 total measurements), measured in a different centre. a) The six selected flow shape characteristics showed moderate to good reliability. b) Site of collapse prediction showed good reliability. CCCp: complete concentric collapse at the level of the palate; LW: lateral wall collapse; TB: tongue base collapse; EG: epiglottis collapse.

External validation

External validation on data from a separate centre (n=466, median (IQR) AHI 28 (19–43) events·h−1, BMI 29 (27–32) kg·m−2, age 57 (48–64) years) showed a significant association between true CCCp and model-predicted CCCp probability score (ncase:ncontrol=14:349; OR 4.0 (95% CI 1.3–12.0) per sd increase). For lateral wall (ncase:ncontrol=84:157), tongue base (ncase:ncontrol=157:131) and epiglottis (ncase:ncontrol=13:275) collapse, similar associations were replicated (OR 2.2 (95% CI 1.3–3.9), 2.4 (95% CI 1.4–4.0) and 6.7 (95% CI 2.5–18.0) per sd, respectively). See “External validation” in the supplementary results for more details.

Discussion

Overall, this study showed that airflow shapes observed in a routine polysomnographic study can provide insight into the likely pharyngeal structures contributing to collapse seen during DISE. As hypothesised, CCCp was characterised by breaths with increased scoopiness. Combining six flow shape characteristics identified a subgroup of patients with 5-fold higher odds of exhibiting CCCp. Flow shape characteristics remained associated with CCCp after adjusting for covariates (AHI, BMI (typically higher in CCCp [18]) plus sex), demonstrating their novel predictive value for site of collapse detection. Furthermore, when assessed in a separate cohort, flow shape characteristics and their predicted site of collapse were repeatable across nights, supporting the notion that these measures can be considered a “trait” and overcome an important prerequisite for future clinical application.

Novel physiological insight

Average flow shapes of patients with CCCp as measured during respiratory events (hypopnoeas) are characterised by increased scoopiness (NED), left-leaning inspiratory flow shapes and several additional features characterising similar behaviour, including faster inspiratory rise time, greater early peak flow (greater initial peak if multiple peaks are present) and a shorter time to peak inspiratory flow as a fraction of the time to peak expiratory flow. On the other hand, a reduced proportion of inspiratory tidal volume occurring in the first 30% of inspiratory time (lower “early inspiratory volume”) was independently predictive of CCCp, which we interpret as a form of calibration or reference for the aforementioned variables. The greater NED seen in CCCp in the current study is consistent with previous work demonstrating higher NED in breaths exhibiting “isolated palate” collapse compared with tongue base collapse (CCCp was not examined in this previous work [11]). Mechanistically, the increased NED seen in CCCp is likely to reflect a greater dynamic reduction and recovery of pharyngeal cross-sectional area within inspiration, given the known association between NED and increased retropalatal compliance [19]. Along these lines, left-leaning breaths are considered to reflect a failure of the airway to respond to a meaningful within-breath increase in pharyngeal muscle activity; in principle, the typical ramp-like increase in pharyngeal muscle activity [20] should promote a gradual rise in inspiratory flow, resulting in right-leaning inspiratory shapes. Consistent with this interpretation, failure of CCCp to respond to increasing muscle activity is a known characteristic of this site/pattern of collapse [2, 21]. The mechanistic bases for the predictive value of the remaining characteristics are unclear.

The current study showed that, like CCCp, patients with lateral wall collapse also exhibited greater scoopiness and inspiratory skewness (left-leaning inspiration) during their respiratory events seen in conventional polysomnography. By contrast, tongue base collapse exhibited diametrically opposite characteristics compared with CCCp or lateral wall collapse (less scoopiness, right-leaning inspiration), consistent with the previous observation of lower NED in tongue base collapse versus other sites [11]. Here, epiglottis collapse characteristics overlapped substantially with tongue base collapse, with the addition of slower inspiratory rise time in epiglottis collapse. Less scoopiness, however, was unexpected given prior findings of greater scoopiness during epiglottis collapse [11, 12]; differences may lie with the use of averaged values here (all hypopnoea breaths) versus individual breath analysis previously [9, 11, 12]. We consider that epiglottis collapse is characteristically intermittent and/or non-sustained and may not always contribute to scored events [22]. A shared aetiology with tongue base collapse, however, is not unexpected: intuitively, a posteriorly located tongue may predispose to epiglottis collapse. These findings, taken together with the exploratory modelling (supplementary tables E14–E16 and supplementary figure E5), also suggested that the primary strength of flow shape characterisation is the ability to discriminate between CCCp/lateral wall collapse, located higher in the airway with a lateral component, and tongue base/epiglottis collapse, located lower in the airway with an anteroposterior component.

Clinical implications

Our study has clear clinical implications for patient selection for non-CPAP treatments, most of which are site-specific interventions. Specifically, CCCp and lateral wall collapse are associated with reduced efficacy of hypoglossal nerve stimulation [2, 6, 23] and mandibular advancement devices [4, 5], while tongue base collapse predicts favourable responses [3, 4]. Our finding that CCCp and lateral wall collapse appear distinct from tongue base collapse in flow shape characteristics therefore provides an avenue for mechanistically informed treatment selection. For example, patients interested in hypoglossal nerve stimulation could be counselled on their preference based on whether they have a high or low CCCp likelihood detected before proceeding with DISE. While preliminary work appears promising for treatment response prediction [24–26], further studies are needed to demonstrate clinical benefit.

Our study makes major steps towards translating the concept that flow shape characteristics differ between sites of collapse into clinical practice. Previous in-lab studies during natural and drug-induced sleep using simultaneous endoscopy and gold standard pneumotachograph flow associated the flow shape of manually selected individual breaths with its concurrent collapse type [9–12]. Here, we showed that automated analysis of the clinical airflow signal from a separate clinical polysomnogram contains the information necessary to predict CCCp likelihood (and other collapse sites, independent of known covariates AHI, BMI [18, 27] plus sex), such that advanced research-level signals data are not necessarily required. It is therefore highly feasible to estimate the probability of different sites of collapse without performing endoscopy.

Methodological considerations

We consider several limitations.

1) For model development, we opted to select patients with distinct presence versus absence of CCCp without the complicating influence of overlapping collapse sites, which involved the exclusion of multi-level collapse (CCCp plus tongue base). We considered whether including these patients could reduce the utility of flow shapes to predict CCCp. Additional analysis (supplementary figure E7) showed that a concurrent complete tongue base collapse masks CCCp to yield flow shape characteristics of non-CCCp. It might thus be difficult to predict CCCp with concomitant tongue base collapse. While this masking could be considered a limitation, tongue base collapse is associated with favourable outcome of several non-CPAP treatments [3, 4], such that the utility for response prediction may not be affected.

2) Model development was done on a rather small final sample size. However, validation of the results in a separate cohort from a different centre could be performed, highlighting the clinical utility of our findings.

3) We used averaged flow shape features to provide representative characteristics for each patient. While previous studies measured airflow with simultaneous endoscopy, the current study employed airflow and DISE data from separate studies. Thus, it was not possible to relate individual breaths to a specific collapse site. For epiglottis collapse in particular, expanding the analysis approach to individual breath-level analyses (e.g. “jaggedness”, “discontinuities” [9]) may allow for greater identification of epiglottis collapse in future.

4) Six flow shape characteristics were selected based on a combination of visual inspection and bivariate analysis with the goal of providing a simple interpretable model. It is possible that other selection techniques might have yielded stronger model performance; additional analyses were performed on each individual model to allow a) additional characteristics or b) replacement characteristics (see “Sensitivity analysis – number of flow shape characteristics used” in the supplementary methods). These models did not yield improvements in performance (cross-validated odds ratios).

5) DISE was used rather than natural sleep endoscopy for the labelled site of collapse, which may provide a source of additional uncertainty [28–30]. We might expect associations to be stronger if natural sleep was used. However, we emphasise that DISE is the current clinical standard used to characterise site of collapse and is known to provide insight into non-CPAP treatment efficacy [2–6, 31].

6) Night-to-night variability affects OSA severity [32] and may be expected to occur regarding flow shape classification. However, analysis of a separate cohort [17] showed good repeatability for individual flow shape characteristics and moderate to good repeatability for site of collapse prediction, suggesting that flow shape-based site of collapse provides a new trait that can be leveraged for better characterisation of the underlying causes of OSA in individual patients.

7) The site of pharyngeal collapse and its associated flow shape could be altered by sleep stage. In the current analysis, we opted to include all breaths during a hypopnoea event, regardless of sleep stage. Sensitivity analysis using only breaths during rapid eye movement (REM) or non-REM sleep suggests the developed model is driven by non-REM breaths (see “Sensitivity analysis – alternative patient and sleep stage criteria” in the supplementary methods and results). Further research is needed to allow evaluation of the site of collapse during REM sleep. The authors argue flow shape analysis could play an important role in this research, as flow shapes can be collected during REM sleep, which is impossible using the classic endoscopy techniques as sleep is too disturbed during natural sleep endoscopy, not allowing the patient's sleep to progress to REM sleep and REM sleep is suppressed during DISE.

8) Finally, we demonstrated that polysomnographic flow shape characteristics could identify patients with increased odds of CCCp (OR ∼5), yet accuracy was modest (73%). Although prediction certainty is not established on an individual level, the approach may help to identify a subgroup of patients for whom CCCp is particularly low likelihood (8% versus 30% in predicted non-CCCp) and thereby help move the field towards more precise clinical intervention.

Conclusions

Overall, the current study found that polysomnographic flow shape characteristics provide insight into CCCp likelihood as observed during a separate DISE procedure. Characteristics were similar between CCCp and lateral wall collapse, and distinctly different from tongue base and epiglottis collapse. By providing a means to predict site of collapse from a routine clinical study, our work has broad implications for the field's goal of judicious provision of efficacious and tolerable therapies for a greater number of patients with OSA.