"CraniOfaCial PHEnOTyPing Of OSa Prediction of Obstructive Sleep Apnea with Craniofacial Photographic Analysis 1,3 2 2 1,3 3 1,3 Richard W. W. Lee, MD ; Peter Petocz, PhD ; Tania Prvan, PhD ; Andrew S. L. Chan, MD ; Ronald R. Grunstein, MD, PhD ; Peter A. Cistulli, MD, PhD 1 2 Centre for Sleep Health and Research, Department of Respiratory Medicine, Royal North Shore Hospital, NSW, Australia; Department of3 Statistics, Macquarie University, NSW, Australia; Woolcock Institute of Medical Research, University of Sydney, NSW, Australia Study Objectives: To develop models based on craniofacial photo- ceiver operating characteristics curve [AUC] 0.82). Combination of pho- graphic analysis for the prediction of obstructive sleep apnea (OSA). tographic and other clinical data improved the prediction (AUC 0.87),Design: Prospective cohort study. whereas prediction based on clinical assessment alone was lowerSetting: Sleep investigation unit in a university teaching hospital. (AUC 0.78). The optimal CART model provided a similar overall classi- Patients: One hundred eighty subjects (95.6% Caucasian) referred for fcation accuracy of 76.7%. Based on this model, 59.4% of the subjectsthe initial investigation of OSA were recruited consecutively. were classifed as either high or low risk with positive predictive valueinterventions: Clinical assessment and frontal-profle craniofacial pho - of 90.9% and negative predictive value of 94.7%, respectively. The re- tographic analyses were performed prior to polysomnography. Predic- maining 40.6% of subjects have intermediate risk of OSA. tion models for determining the presence of OSA (apnea-hypopnea in- Conclusions: Craniofacial photographic analysis provides detaileddex [AHI] = 10) were developed using logistic regression analysis and anatomical data useful in the prediction of OSA. This method allowsclassifcation and regression trees (CART). OSA risk stratifcation by craniofacial morphological phenotypes. Measurements and results: Obstructive sleep apnea was present Keywords: Obstructive sleep apnea; craniofacial abnormalities; pho- in 63.3% of subjects. Using logistic regression, a model with 4 pho- togrammetry; prediction tographic measurements (face width, eye width, cervicomental angle, Citation: Lee RWW; Petocz P; Prvan T; Chan ASL; Grunstein RR; Cis- and mandibular length 1) correctly classifed 76.1% of subjects with and tulli PA. Prediction of obstructive sleep apnea with craniofacial photo- without OSA (sensitivity 86.0%, specifcity 59.1%, area under the re - graphic analysis. SLEEP 2009;32(1):46-52. OBSTRUCTIVE SLEEP APNEA (OSA) IS A VERY COMMON We have developed a photographic analysis technique whichDISORDER ASSOCIATED WITH SNORING, REPETITIVE allows detailed quantitative assessment of craniofacial mor- UPPER AIRWAY COLLAPSE DURING SLEEP, OXYGEN de- phology. These craniofacial photographic measurements appear1,2 saturation and sleep fragmentation. It is associated with increased to capture a number of risk factors relevant to OSA, includingcardiovascular morbidity, motor vehicle accident risk, and overall skeletal restriction, regional adiposity and obesity (see com- 3 mortality. The diagnosis of OSA is cumbersome because of the panion article “Craniofacial Phenotyping in Obstructive Sleep15 need for specialist assessment and overnight monitoring in a sleep Apnea - A Novel Quantitative Photographic Approach” ). Thislaboratory. The latter is expensive, labor intensive, and resource technique could be useful in a number of clinical and research4 limited. As a result, the recognition of OSA in the community is applications where high throughput is a requirement, such as5 low, and the majority of sufferers of OSA are as yet undiagnosed. in epidemiological research. We hypothesized that craniofacialHence, there is a critical clinical need to develop methods to im- photographic analysis would also be a useful technique in theprove recognition and diagnosis of OSA in the community. prediction of OSA. The primary aim of this study was to de- Prediction algorithms have been developed for risk stratifca - velop prediction models based on craniofacial photographiction and screening of subjects for OSA. These algorithms are measurements for the prediction of OSA, and to compare thesebased mainly on data such as patient demographics, symptoms, to models based on other clinical data. 6,7 and measures of obesity. While obesity is generally consid- 8 ered the major risk factor for OSA, craniofacial morphology METHODS is increasingly recognized as an important interacting factor9-11 in OSA pathogenesis. However, craniofacial or intraoral Subjects risk factors are included in only a minority of OSA prediction12-14 algorithms. This relates to the impractical nature of the cur- Subjects referred for polysomnography to a university teachingrently available craniofacial assessment techniques. Further- hospital for the initial investigation of OSA were recruited con- more, the suboptimal accuracy of these clinical algorithms and secutively. Exclusion criteria included those with a known historycomplexity of some measurement techniques limit their routine of syndromal craniofacial abnormalities (e.g., Down syndrome),use in the clinical diagnosis of OSA. previous craniofacial surgery, and excessive facial hair that sig- nifcantly obscured facial landmarks. Subjects of all ethnicity(self-reported) were included. Clinical assessment and the stan- Submitted for publication June, 2008 dardized photographic procedure were performed on all subjectsSubmitted in fnal revised form August, 2008 on the same day as the polysomnography. All data collection andaccepted for publication august, 2008 photographic analyses were carried out by a single investigatorAddress correspondence to: Peter Cistulli, MD, PhD, Centre for Sleep(RL) who was blinded to the result of polysomnography. EthicsHealth and Research, Department of Respiratory Medicine, Level 8, Mainapproval was obtained from the institutional ethics committee,Block, Pacifc Highway, St Leonards, NSW 2065, Australia; Tel: +61 2and written informed consent was obtained from all subjects. 9926 8674; Fax: +61 2 9906 6391; E-mail: [email protected] SLEEP, Vol. 32, No. 1, 2009 OSA Prediction with Craniofacial Photographs—Lee et al 46Standardized Photographic Technique Polysomnography Frontal and profle photographs of the head and neck were Diagnostic polysomnography (PSG) was performed in ac- 17,18 obtained with a standardized setup using a single-lens refex cordance with previous studies and recommendations. Sleep19 digital camera (D70 with 18-70mm lens and external fash unit staging was determined using standardized defnitions. ApneaSB-29s; Nikon Corp., Japan). Prior to the photographs, certain was defned as complete airfow cessation for = 10 seconds withbony and cartilaginous landmarks were pre-identifed on the oxygen desaturation of at least 3% and/or associated with arous- subjects by palpation and marked with a white tape. The stan- al. Hypopnea was defned as a reduction in amplitude of airfowdardized technique used for subject alignment and its test-retest or thoracoabdominal wall movement > 50% of the baseline mea- reliability are described in the companion article “Craniofacial surement for > 10 seconds with an accompanying oxygen desatu- Phenotyping in Obstructive Sleep Apnea - A Novel Quantitative ration of at least 3%, and/or associated with arousals. Apnea-hy- 15 Photographic Approach.” popnea index (AHI) was calculated as the total number of apneasand hypopneas per hour of sleep. Polysomnography scoring wasCraniofacial Photogrammetry performed by experienced accredited sleep technologists. Using image analysis software (Image J v1.36, NIH, Bethes- Data and Statistical analysis da, MD), the photographs were examined for landmark digitiza- tion. Craniofacial landmarks of interest were captured as pixel Predictive models for OSA were developed using 2 differentcoordinates (x, y) of the image which were then transferred to statistical approaches, namely logistic regression (SPSS v13.0a custom-programmed spreadsheet for the computation of lin- for Windows, SPSS Inc., Chicago, IL, USA) and classifcationear, angular, area, and polyhedral volume measurements. Pixel and regression tree (CART) analyses (Salford Systems [2006],measurements were converted to metric dimensions based on CART Extended Edition Version 6.0, San Diego, California,20 a conversion scale of 52 pixels/cm. In addition to the 71 mea- USA). In both analyses, the presence of OSA was defned bysurements obtained in the previous study, another 62 related an AHI = 10 events per hour and those without OSA were de- craniofacial measurements were included (133 measurements fned by an AHI of < 10 events per hour. in total). These measurements represented the dimensions andrelationships of the various craniofacial regions including the logistic regression face, mandible, maxilla, eyes, nose, head, and neck (see supple- mentary data: Appendix 1). All 133 photographic measurements were initially consideredand a multistep process was employed to reduce the number ofClinical assessment variables for further analysis. Multi-colinearity reduced the totalnumber of measurements to 105. These were further reduced us- Subject data on demographics, symptoms of OSA, comor- ing forward stepwise regression of the log-transformed AHI (withbidities, and Epworth Sleepiness Scale (ESS) were obtained the addition of 1) for each group of measurements (linear, angles,by questionnaire. Anthropometric assessment included neck areas, and volumes). This approach led to the reduced set of 13circumference, waist circumference, and body mass index variables (see Appendix 1: L27, L61, L62, L65, AN18, AN19,(BMI). Oropharyngeal assessment was performed with stan- AR3, AR9, AR14, AR20, V2, V13, and V19) for further analysis12,14,16 dardized techniques as described in previous studies. to derive the OSA prediction models. Forward likelihood ratio lo- These included the assessment of the modifed Mallampati gistic regression of the remaining set of variables was employedclass (MMC) (assessed with mouth wide open without pro- to generate the photographic prediction model for OSA (Logistictrusion of the tongue: [I] tonsils, pillars and soft palate were Regression Model 1). Logistic Regression Model 2 was developedclearly visible; [II] uvula, pillars and upper pole were visible; by replacing selected variables from Logistic Regression Model 1.[III] only part of the soft palate was visible; [IV] only the hard Backward likelihood ratio logistic regression was used to developpalate was visible), pharyngeal grade (assessed with mouth the clinical and combined clinical/photographic prediction mod- wide open and maximal protrusion of the tongue: [I] palatopha- els (Logistic Regression Models 3 and 4). A total of 16 clinicalryngeal arch [ppa] intersects at the edge of the tongue; [II] variables were considered (age, sex, BMI, neck circumference,ppa intersects at 25% or more of the tongue diameter; [III] waist circumference, hypertension, diabetes mellitus, alcohol useintersects at 50% or more; [IV] intersects at 75% or more), = 20g/day, witnessed apneas, ESS, MMC, pharyngeal grade, ton- tonsillar grade ([I] previous tonsillectomy or tonsils not seen; sillar grade, enlarged uvula, enlarged tongue, and overjet) for the[II] tonsils visible behind the anterior pillars; [III] tonsils ex- clinical models. Classifcation accuracy, model characteristics,tended 75% of the way to the midline; [IV] tonsils completely predictive values, and receiver operating characteristic (ROC)obstructing airway), uvula size (considered enlarged if its ap- curves were calculated for each model. The probability thresholdproximate length is > 1.5 cm and width > 1 cm), tongue size used for classifcation of OSA was 0.50. (considered enlarged if its superior border was above the levelof the mandibular occlusal plane, in association with tongue Classification and regression Tree (CarT) ridging) and the presence of overjet (present if there was agreater than 3 mm anterior-posterior distance between the up- Classifcation and regression tree analysis is a predictive methodper and lower incisors during occlusion). that uses nonparametric techniques to evaluate data and account for21 complex relationships. In this type of analysis, there is progres- SLEEP, Vol. 32, No. 1, 2009 OSA Prediction with Craniofacial Photographs—Lee et al 47Table 1—Subject Characteristics N (%) Mean ± SD Range Number of subjects 180 - - Males (%) 137 (76.1%) - - Age (years)53.4 ± 14.3 20–86 Ethnicity – Caucasians 172 (95.6%) - - Anthropometry 2BMI (kg/m ) - 29.3 ± 5.13 19.5–50.9Neck circumference (cm) - 41.3 ± 4.48 30.5–55.0Waist circumference (cm) - 105 ± 14.4 70.0–148 SymptomsEpworth Sleepiness Scale (ESS) - 8.82 ± 4.99 0–23Witnessed apneas 86 (47.8%) - - PolysomnographyTotal AHI - 22.7 ± 21.7 0–110Minimum SaO (%) - 83.5 ± 10.1 36–99 2 -z sive splitting of the population into subgroups that are based on the = 10) can be calculated by the formula: 1 / (1+e ), where z =predictive independent variables. The variables chosen, discrimi- -9.235 + 1.442 (face width [cm]) - 2.872 (eye width [cm])natory values of the variables, and the order in which the splitting + 0.02 (cervicomental angle [degree]) - 1.224 (mandibularoccurs are all produced by the underlying mathematical algorithm length 1 [cm]). This model classifed 76.1% of the subjectsto maximize predictive accuracy. A 10-fold cross-validation pro- correctly. It had a sensitivity of 86.0%, specifcity of 59.1%,cess was applied during the development of the CART models in positive predictive value (PPV) of 78.4% and negative predic- order to minimize over-ftting of the data. This cross-validation tive value (NPV) of 70.9%. The area under the ROC curve wasprocedure involved modelling using a proportion (90%) of the data 0.82 (Figure 1A). In those who were incorrectly classifed, theand validation with the remaining (10%), and then repeating with mean AHI was signifcantly lower than those who were cor - a different one-tenth of the data until all data have been covered. rectly classifed (12.9 versus 25.9 events/hr, P < 0.001); butAll 133 photographic measurements and every value of splits of there was no difference in age, BMI or neck circumferencethese measurements were analyzed with CART in order to con- between these groups. struct models that can optimally separate subjects with and withoutOSA. The classifcation trees were built by continuing splitting of logistic regression Model 2 – Uncalibrated Photographiccases to achieve “terminal nodes” which are clusters of cases with Measurements or without OSA. Models using a single photographic measure- ment, multiple photographic measurements, and combinations of In order to allow uncalibrated photographs to be used forphotographic and clinical measurements were constructed. OSA prediction, Logistic Regression Model 1 was simplifedby using ratio and angular measurements instead of calibratedrESUl TS metric measurements. This model used the face width-eyewidth ratio to replace the individual measurements and man- Subject Characteristics dibular-nasion angle 2 instead of mandibular length 1. Prob- ability of OSA can be calculated using z = -4.516 + 1.528A total of 180 subjects were included in the analysis; ob- (FER [face width-eye width ratio]) + 0.025 (cervicomentalstructive sleep apnea (AHI = 10) was present in 114 subjects angle [degree]) - 0.262 (mandibular-nasion angle 2 [degree]).(63.3%). Three subjects were excluded from analysis (2 subjects This model classifed 71.1% of the subjects correctly. It haddid not complete the PSG; 1 subject was found to have central a sensitivity of 80.7%, specifcity of 54.5%, PPV of 75.4%,sleep apnea). Twelve subjects (5 females, 7 males) declined and NPV of 62.1%. The area under the ROC curve was 0.80study participation. Characteristics of the subjects, clinical data, (Figure 1A). and polysomnographic indices are summarized in Table 1 andAppendix 2 (supplementary data). Details of the logistic regres- logistic regression Model 3 – Clinical Measurements sion models are contained in Appendix 3 (Supplementary dataavailable at www.journalsleep.org). This prediction model for OSA was built using all the clini- cal variables. Age, BMI, and witnessed apneas were identifedlogistic regression analysis as independent predictors for OSA. This model classifed 76.1%of the subjects correctly. It had a sensitivity of 86.0%, specifc - logistic regression Model 1 – Calibrated Photographic ity of 59.1%, PPV of 78.4%, and NPV of 70.9%. The area underMeasurements the ROC curve was 0.78 (Figure 1B), which was smaller thaneither of the photographic models (Logistic Regression ModelsThis model had the highest overall correct classifcation 1 and 2). with the least number of variables. Probability of OSA (AHISLEEP, Vol. 32, No. 1, 2009 OSA Prediction with Craniofacial Photographs—Lee et al 48A – Logistic Regression Models 1 and 2Figure 2—CART Model 1: Single Photographic Measurement.Fifty-fve out of 60 (91.7%) subjects in terminal node 2 (*) hadOSA. (AN19 = mandibular width-length angle [degrees]. Class0 = No OSA; Class 1 = OSA.) for OSA. Witnessed apnea and MMC were the only clinicalB – Logistic Regression Models 3 and 4variables further contributing to the model, although the con- tribution of MMC was small. This combined photographic andclinical model classifed 79.4% of the subjects correctly. It hada sensitivity of 85.1%, specifcity of 69.7%, PPV of 82.9%, andNPV of 73.0%. The area under the ROC curve was highest at0.87 (Figure 1B). Classification and regression Tree (CarT) analysis CarT Model 1 – Single Photographic Measurement The simplest CART model used a single photographic mea- surement (mandibular width-length angle) to classify 64.4% ofthe subjects correctly with 2 terminal nodes (Figure 2). In thismodel, if the mandibular width-length angle was > 89.8 de- grees, 55 out of 60 subjects (91.7%) had OSA. If the angle was= 89.8 degrees, 61 of 120 subjects (50.8%) did not have OSA.This model had a sensitivity of 48.2%, specifcity of 92.4%,PPV of 91.7%, and NPV of 50.8%. In other words, one-third ofthe cases (60 of 180 subjects) were classifed as having a highrisk of OSA of 91.7%. Figure 1—Receiver Operating Characteristic (ROC) Curves forLogistic Regression Models. A – Logistic Regression Models 1CarT Model 2 – Multiple Photographic Measurements (Calibrated Photographic Measurements) and 2 (UncalibratedPhotographic Measurements); B – Logistic Regression Models 3This model used 4 photographic measurements (mandibular(Clinical Measurements) and 4 (Photographic and Clinical Mea- width-length angle, neck depth, mandible width, face width-lowersurements). AUC = area under the curve. face depth angle) to classify 76.7% of the subjects correctly with 5terminal nodes (Figure 3). This model had a sensitivity of 70.2%,logistic regression Model 4 – Photographic and Clinical specifcity of 87.9%, PPV of 90.9%, and NPV of 63.0%. Based onMeasurements this model, 80 of 88 subjects (90.9%) in terminal nodes 2, 4, and5 collectively had OSA and 18 of 19 subjects (94.7%) in terminalThis model was developed with the reduced set of 13 photo- node 1 did not have OSA. The remaining 73 subjects at terminalgraphic measurements and all the clinical variables. Similar to node 3 had intermediate risk of OSA of 45.2%. In other words,Logistic Regression Model 1, face width, eye width, and man- 59.4% (107 of 180) of the subjects were classifed as either high ordibular length 1 remained independent photographic predictors low risk with PPV of 90.9% and NPV of 94.7%, respectively. SLEEP, Vol. 32, No. 1, 2009 OSA Prediction with Craniofacial Photographs—Lee et al 49"