Predicting sleep apnea from three-dimensional face photography

INTRODUCTION

Obstructive sleep apnea (OSA) is a common disorder estimated to affect 15% of middle‐aged males and 5% of middle‐aged females.¹ It is characterized by repetitive episodes of partial or complete upper airway obstruction that are associated with hypoxemia, sympathetic activation, and sleep disruption. It is associated with burdensome symptoms and substantial medical comorbidities, including sleepiness-related accidents,² diabetes, cardiovascular diseases,³ and depression.⁴

Studies demonstrate familial aggregations of OSA, suggesting powerful genetic predispositions, although these remain to be precisely characterized. Despite OSA being readily treatable, most cases (75%) remain unidentified,^5,6 because symptoms and signs are unrecognized, ignored, or misattributed to other causes. Current screening tools involve questionnaires, which while reasonably sensitive for OSA are (relatively) nonspecific, resulting in a high rate of false-positive results.⁷ Other low-cost screening tools are needed.^8–11

Craniofacial anatomy is recognized as an important predisposing factor in OSA pathogenesis.¹² Studies using magnetic resonance imaging (MRI) have shown that midface and lower-face width are correlated with OSA severity, supporting the notion that facial structure is important in the development of OSA.¹³ Lateral cephalometry (x-ray) studies have demonstrated the importance of facial measures such as maxillary and mandibular length and intermaxillary space¹⁴ in determining risk of OSA. Cephalometry and MRI allow accurate measurement of specific dimensions of the facial skeleton and upper airway. However, these techniques are not available for routine clinical assessment and their use is limited by cost in the case of MRI and the risks associated with exposure to ionizing radiation in the case of cephalometry.

Two-dimensional (2D) photography has been used as an alternative to these more complex techniques and has several advantages. In particular, it is safe, inexpensive, portable, and easily accessible. Facial phenotypes assessed with 2D photography are closely correlated with upper airway anatomy as determined using MRI. Hence, 2D photography represents a reliable method for assessing both internal and external facial structures.¹³ 2D photography captures several anatomic risk factors for OSA related to skeletal restriction, regional adiposity, and obesity. Such craniofacial risk factors include a wider and flatter mid and lower face, a shorter and retruded mandible, a smaller enclosed area within the mandible, and more soft tissues or fat deposition on the anterior neck.^12,15 Measurements such as these have been used to predict OSA severity with reasonable accuracy.^15,16 However, 2D photography cannot capture the nonlinear nature of craniofacial anatomy, such as shape and contour.

Three-dimensional (3D) photography overcomes this limitation of representing the 3D structure of the face with 2D imaging. Although similar to 2D photography it can provide information regarding linear distances and angles, unlike 2D photography it can additionally provide information on facial contours (ie, geodesic distances) making it an ideal tool for assessing the role of craniofacial structure in the pathogenesis of OSA. 3D photography allows assessment of the skeletal and soft tissues of the face and neck in a faster, cheaper, more readily available, and less invasive manner than MRI¹⁷ and is already used in applications that range from measuring aesthetic facial parameters¹⁸ to orthodontic diagnosis and evaluation of the craniofacial growth and development.¹⁹ To date there has been only one study examining the potential role of 3D facial analysis to predict OSA in adults. In this study, Lin et al²⁰ showed in 36 male Asian patients with OSA that linear measurements of craniofacial distances, areas, angles, and volumes captured by 3D photography showed strong agreement with the same measurements obtained with 3D computed tomography (CT).

The current study aimed to use 3D facial surface analysis of linear and geodesic measurements to determine the combination of measurements that best correlates with the presence and severity of OSA. In particular, we sought to determine whether geodesic measurements increased the accuracy with which individuals with OSA could be identified from 3D photographs of the face.

METHODS

Landmark annotation

A total of 24 landmarks were annotated by a single scorer who was blinded to OSA status, on the 3D image of each face (Figure 1) using custom software developed by the authors (AM, SG) in Matlab (R2018b) which allowed these landmarks to be objectively positioned. Two of these landmarks (left and right gonion) were physically marked on each individual’s face at the time of scanning. Selection of landmarks were guided by earlier studies of Sutherland and Lee.^15,26,27 The software allowed the face to be rotated to allow optimal visualization and determination of each landmark. Landmark location was highly repeatable: the average intraclass correlation coefficient for the 22 landmarks was 0.99 when calculated from 20 faces (5 from each OSA category annotated twice each), excluding the gonions as they were physically marked on the face. The 3D coordinates of each landmark (in the x, y, and z axes) were stored for later analysis.

Feature extraction

The 24 landmarks allowed a total of 276 distances between two (paired) landmarks to be defined. For parsimony and guided by earlier studies^12,26,27 25 pairs of points were selected and distances calculated for each in both the linear dimension (ie, direct euclidian distance between the two points) and geodesic dimension (ie, the shortest distance between two points when following the contour of the face/skin) (Figure 2, Table S1 and Table S2 in the supplemental material). Using linear dimensions only, angles were also determined between sets of three points (Figure S1 and Table S3 in the supplemental material). These linear and geodesic distances and angles were termed “features” and used in further analysis.

RESULTS

Data from 400 participants (172 males), all of whom had 3D face scans, overnight laboratory-based sleep studies, and measurements of neck circumference and BMI, were used in this study to develop, train, and validate the predictive algorithm (Table 1 and supplemental text). Ethnicity data were available on 155 of the participants of whom 72% were classified as Caucasian (both parents were Caucasian). Compared to those without OSA (AHI < 5 events/h), those with OSA (AHI ≥ 5 events/h) were older (47.4 ± 13.8 versus 54.5 ± 15.8 years, P < .05) and had increased AHI (2.6 ± 1.28 versus 30.7 ± 26.0 events/h, P < .05), BMI (26.9 ± 5.2 versus 32.3 ± 7.6 kg/m², P < .05) and neck circumference (35.0 ± 3.8 versus 39.7 ± 4.8 cm, P < .05) (Table 2).

The accuracy with which facial measurements, BMI, and neck circumference could classify an individual as having OSA (AHI ≥ 5 events/h) or not (AHI < 5 events/h) is summarized in Table 3 and shown graphically in Figure 3 (the algorithm is reported in the supplemental text). The predictive accuracy of geodesic measurements alone was significantly greater than that for linear distances alone, being 89 ± 1% and 86 ± 1%, respectively (P < .01). In general, the accuracy to predict the presence of OSA was lowest for simple anthropometric measures of BMI and neck circumference, was increased for measurements of linear or geodesic distances, and was maximal (with an accuracy of 91 ± 2%) for measurements incorporating a combination of linear and geodesic distances. The classification accuracy to predict the presence of OSA was not significantly different whether considering females alone or males alone (Table 4).

Differences in landmark distances between those with and without OSA (based on a cutoff of 5 events/h) are shown for geodesic distances, linear distances and angles in Table S1, Table S2, and Table S3. Craniofacial features that differed between the control and OSA groups, for both geodesic and linear distances included: upper, mid, and lower face depth; face width; mandibular length, width and posterior height; and neck width (P < .05 for all). The most discriminatory angular measurements between the two groups (based on magnitude of difference and values of P) were the mandibular width angle (right gonion-menton-left gonion); lower facial width angle (right tragion-menton- left tragion); and maxillary-mandibular relationship angle (subnasion-nasion-gnathion) (P < .05 for all).

Several of these geodesic and linear distances were different between control and mild, moderate, and severe OSA and between each of the different severities of OSA (Table S4 and Table S5 in the supplemental material). These included: upper and lower face depth; total and upper face height; lateral face height; face width; mandible width; and neck width (P < .05 for all). Angular measurements tended to be less discriminatory between controls and different severities of OSA (Table S6 in the supplemental material).

The classification accuracy for predicting OSA was related to the AHI cutoff used to define the presence or absence of OSA. This is shown in Figure 4, which demonstrates that, relative to a cutoff of 5 events/h, accuracy decreases when using the two other most commonly used cutoffs of 10 and 15 events/h (note that the algorithm was retrained at each cutoff). However, regardless of the cutoff used the highest accuracy was achieved by using a combination of linear and geodesic distances.

DISCUSSION

The study was driven by the need to develop new, effective screening methods for OSA, given the high prevalence combined with low identification; the disappointingly low specificity of existing screening techniques (questionnaires) and the prospect that 3D craniofacial photography could improve that; and that 3D photography techniques are widely available with more reliably derived, less time-consuming measures than 2D facial photography, which requires formalized anteroposterior and lateral photographs. The current study showed that, relative to linear or anthropometric measurements, geodesic (nonlinear) measurements of craniofacial anatomy improved the ability to identify individuals with and without OSA. Maximal classification accuracy (91%) was found when geodesic and linear craniofacial measurements were combined into a single predictive algorithm. Such screening could represent the first inexpensive, widely available step along the diagnostic pathway for OSA.

The notion that surface facial dimensions can identify individuals with OSA is supported by findings from previous studies showing that dimensions of the facial skeleton and soft tissues are related to upper airway dimensions^12,13 with airway narrowing an important predisposition to its collapse during sleep. Furthermore, many studies across different ethnic populations have shown that measurements of surface craniofacial dimensions are correlated with OSA severity.^{12–15,26,29–31} The techniques used in these previous studies have included MRI, CT, lateral cephalometry, and 2D photography. Although cephalometry and MRI allow accurate measurement of specific dimensions of the facial skeleton and upper airway, these techniques are not suitable for routine clinical assessment and their use is limited by cost in the case of MRI and the risks associated with exposure to ionizing radiation in the case of cephalometry and CT. Craniofacial measurements from 2D photography have been used to predict OSA severity with reasonable accuracy^15,16; however 2D photography is unable to effectively capture the nonlinear nature of craniofacial anatomy, such as shape and contour.

3D photography overcomes the limitation of representing the 3D structure of the face with 2D imaging because it has the ability to provide information regarding facial contours as well as the linear distances and angles available from 2D images. Furthermore, compared to 2D photography, 3D measurements of craniofacial distances and angles (obtained using the 3dMD system) have increased agreement with equivalent measurements obtained from 3D CT.²⁰ The current study showed that the accuracy with which OSA can be predicted from craniofacial photographs is increased when information on facial contours is considered, regardless of the AHI threshold used to define the presence of OSA. Specifically, geodesic measurements alone were more accurate than linear measurements alone, with maximum accuracy derived from a combination of geodesic and linear measurements. The sensitivity, specificity, accuracy, and area under the ROC curve of the algorithm to predict OSA (AHI ≥ 5 events/h) based on this combination was 97%, 76%, 91%, and 0.96, respectively. Notably, these values were similar when considering the group overall, or males and females separately.

These measures of predictive capacity tend to be higher than previous studies using 2D facial photographic analysis, which have shown that correct OSA risk classification can be obtained in 76% to 79% of the cases, with sensitivities ranging from 73% to 85%, specificities from 28% to 70%, and area under the ROC curve from 0.73 to 0.87.^15,26,32 When comparing between studies and between techniques an important consideration is the AHI cutoff used to define the presence or absence of OSA. In all of these previous studies a cutoff of 10 events/h was used, whereas the current study used a cutoff of 5 events/h. It was interesting to note that, in the current study, the sensitivity, specificity, accuracy, and area under the ROC curve decreased to 90%, 77%, 87%, and 0.92, respectively, when an AHI threshold of 10 events/h was used, and to 89%, 82%, 86%, and 0.92, respectively, when an AHI threshold of 15 events/h was used, despite the algorithm being trained and then tested to classify new unseen cases at each threshold. This indicates that an AHI threshold of 5 events/h is where the accompanying facial morphology changes are maximal. It was notable, however, that regardless of the threshold used the maximum capacity to identify individuals with OSA was obtained when geodesic and linear craniofacial measurements are combined into a single predictive algorithm.

Craniofacial features that differed between the control and OSA groups tended to change similarly for both geodesic and linear distances, such that individuals with OSA were characterized by a 3% to 10% increase in measures of face depth; 10% to 13% increase in lateral face height; 5% increase in face width, 2% to 6% increase in mandibular length; 17% to 19% increase in posterior mandibular height; 4% to 5% increase in mandible width; and 10% to 13% increase in neck width. In terms of angular measurements, individuals with OSA had an 18% smaller maxillary-mandibular relationship angle, and a 10% to 12% decrease in angle between the menton and left/right mandible or tragion. It is difficult to compare these differences to those reported in previous studies using 2D measurements, as the distances and angles derived in the present study were calculated between points in 3D space (ie, each point having an x, y, and z coordinate), and are hence a composite of a 2D frontal view and a 2D lateral view. The only study using 3D craniofacial photography with which to compare our results was published recently by Lin et al.²⁰ Their analyses did not include geodesic distances; however, the linear measurements they found to be correlated with severity of OSA (eg, mandibular width, neck perimeter, mandibular length, facial width, binocular width) were similar to those that distinguished groups in the current study.

The data presented in this study suggest that 3D craniofacial photography has a potential role in screening for OSA in the general population, but not its diagnosis. Notably, even though the accuracy and sensitivity of the algorithm based on the combination of linear and geodesic distances were high at 91% and 97%, respectively, the specificity was 76% and the positive likelihood ratio (= sensitivity / (1 − specificity)) was only 4.04. These findings are consistent with a good screening test, which must have a high sensitivity (so that most cases are identified) and at least moderate specificity to ensure against too many false-positive results. However, they are inadequate for a diagnostic test that must be both highly sensitive and highly specific. In their review of nonlaboratory based devices to diagnose OSA, using an in-laboratory cutoff of 5 events/h, Collop et al³³ suggested that a useful diagnostic test should have a positive likelihood ratio > 5 and sensitivity of at least 0.825. However, it remains possible that its application in a population with increased pretest probability based on clinical assessment³⁴ could result in greater diagnostic accuracy, as might further refinements to the algorithm.

A further finding in this study was that there were systematic differences in several of the geodesic and linear distances between patients in the control, mild, moderate and severe OSA groups, although angular measurements tended to be less discriminatory. This finding suggests that, beyond identification of the presence of OSA, 3D photography may be helpful in predicting OSA severity. Although the internal validity of the data has been assessed in the methods used here, external validation in an independent sample of participants would be a desirable next step in evaluating this method for clinical use.

The method described in this study demonstrates the potential for 3D facial photographs to provide a rapid, simple, objective, and accurate method to identify individuals at high risk of OSA. Although it was developed using an expensive, high-resolution, research quality craniofacial scanner system, it might be possible to embed such a method into existing, cheap, off-the-shelf 3D photography systems. This could potentially provide a novel, accurate screening tool for widespread use in the general population to identify individuals at high risk of OSA. Such a screening tool is desirable given the high prevalence of OSA and its current low identification rates^5,35 and could represent a first, inexpensive, widely available step along the diagnostic pathway.

This study has several limitations. First, individuals with beards were excluded. This was necessary to obtain a full set of anatomic landmarks on all participants. However, beards are often present in those with maxillomandibular deficiency; thus, exclusion may skew the population. Second, the conclusions are limited to a middle-aged population. However, the prevalence of OSA is greatest and therefore the need for such screening tools is also greatest in this population. Third, the conclusions are limited to Caucasians. It is highly likely that different combinations of dimensions will be required to predict OSA in other ethnic groups given the well-known effect of ethnicity on craniofacial phenotype and OSA anatomic risk factors.^36,37 Finally, the patients with OSA had a higher BMI, were older, and had a larger neck circumference than those without OSA. These features may affect facial parameters measured by 3D photography and could confound the results. However, as a real-world screening test our study shows that 3D photography was more predictive than BMI alone (Figure 3) in distinguishing patients with OSA from those without OSA.

In conclusion, the main findings of this study were that geodesic measurements add value to the capacity to identify patients with OSA from 3D photographs of the face and that a combination of linear and geodesic measures has a strong predictive value for the presence of OSA in the studied population.

DISCLOSURE STATEMENT

All authors have seen and approved this manuscript. Work for this study was performed at the University of Western Australia. The Raine Study was supported by the National Health and Medical Research Council, with additional funding for core management provided by the University of Western Australia, Raine Medical Research Foundation, Telethon Kids Institute, University of Western Australia Faculty of Medicine, Dentistry, and Health Sciences, Women and Infants Research Foundation, Curtin University, The University of Notre Dame, Edith Cowan University and Murdoch University. The 22-year follow-up was supported by the National Health and Medical Research Council (1021858, 1027449, and 1044840). Peter Eastwood was supported by a National Health and Medical Research Council Senior Research Fellowship (No. 1042341). The authors report no conflicts of interest.

REFERENCES