" Apnoea-hypopnoea Index Estimation usingCraniofacial Photographic MeasurementsHadis Nosrati, Nadi Sadr, Philip de ChazalCharles Perkins Centre, School of Electrical and Information Engineering,Faculty of Engineering and Information TechnologiesThe University of Sydney, Sydney, Australia collapsibility which can result in OSA. Hence, in majorityAbstractof patients, a combination of imaging methods illustratingcraniofacial abnormalities and measures of obesity couldThis paper presents a novel way of estimating thebe utilized as a tool for OSA recognition and measuringapnoea-hypopnoea index (AHI) using craniofacialthe OSA severity. The current available craniofacialphotographs.We compared the correlation andevaluation tests and imaging including cephalometry,classification performance of the photograph-determinedcomputed tomography (CT) and magnetic resonanceAHI against expert-determined AHI for a number offiltering (MRI) are expensive and invasive. Therefore,selected measurement sets. Our best performing systemphotogrammetry has been widely used as an interestingused five craniofacial measurements selected from 71OSA diagnosis tool due to its non-invasive, accessible andmanual craniofacial phenotype features, which had beenquantifiable nature. Recent studies investigateddetermined from frontal and profile photographs of aquantitative photographic analysis of the craniofacialpatient’s head and neck. The measurements weremorphological phenotype of OSA patients and normalprocessed with a Support Vector Machine Regressionsubjects [2]–[4]. A few craniofacial photographicalgorithm to estimate AHI. The best features includedmeasurements have been discovered with the ability offace width, mandibular length, binocular width, cranial detecting OSA [3].base area, and criocomental space distance. A databaseThis study looks at the craniofacial photographicof 114 subjects with OSA (AHI 10/h) and 66 controlsmeasurements, as a powerful and non-invasive tool to(AHI <10/h) was used for algorithm development andpredict OSA severity. It uses the estimated apnoea- testing. Leave-one-record-out cross-validation was usedhypopnoea index (AHI) as a surrogate measure to classifyto estimate performance. The Pearson correlation wassubjects with and without OSA. The benefits of0.52 for the AHI estimation. Classification performedphotogrammetry for OSA diagnosis over the standard in- using an AHI threshold of 10 events per hour, resulted inlab overnight polysomnogram test include: independencean estimated accuracy of the algorithm of 73.3% with anof time of day for the administering of the test, labourarea under the ROC of 0.78. inexpensive, cost effective and minimally invasive. Thesebenefits could also lead to its application in developingcountries where expensive diagnostic tools may not be1. Introduction readily available.Obstructive Sleep Apnoea (OSA) is a widespread sleeprelated respiratory disorder involving consecutive2. Database blockage of the upper airway. Although, it can lead toserious health issues such as cardiovascular disorders,The landmark features used in our study are selectedabout 80% of OSA cases are undiagnosed [1]. The currentfrom manual photographic measurements representing theOSA diagnosis tools are invasive and expensive. Thus,dimension of various craniofacial regions including face,there is a trend toward developing reliable non-invasivehead, neck, eyes, nose, mandible and maxilla as describedOSA detection methods. in Table 1 and Fig. 1. Measurements are derived from theIt has been reported in the previous imaging studiesstudy conducted by Lee et.al [2] where subjects werethat anatomical and functional abnormalities of the upperreferred for polysomnography to a university teachingairway and craniofacial morphology could be significanthospital for the initial investigation of OSA. A total of 180factors in OSA detection. It has been shown thatsubjects were included in the analysis where 114 subjectscraniofacial abnormalities can lead to upper airwayComputing in Cardiology 2016; VOL 43 ISSN: 2325-887XDOI:10.22489/CinC.2016.297-381Table 1. Examples of71 Craniofacialphotogrammetry features [2]Feature LandmarksBiocular width exl-exrCervicomental angle np-cer-meCranial base area 1(ax) tl-exl-exr-trCricomental space distance cer-cr-meEye width exl-enlFace width tl-trMandibular length 2 gn-goMandibular nasion angle 2 go-n-meFigure 1. Photographic Landmarks – Profile and Frontal Viewhad OSA (AHI>=10/h) and 66 were selected as controls to map the input features to an estimated AHI. We(AHI <10/h) [2]. describe SVR in the section 3.2. Before the photographs were taken, certain landmarks Sections 4 and 5 provide the details of the experiments,were identified on the subjects and indicated with a white the results, and the conclusions. tape. A calibration washer of known size was taped to theforehead and to the cheek. It was used to calibrate3.1. Craniofacial Photographic Featuresmeasurements determined from the photograph. Frontaland profile photographs of the head and neck were thenobtained with a single-lens reflex digital camera (D70 with Three systems were considered as the photographic18-70mm lens and external flash unit SB-29s; Nikonfeatures for AHI estimation. Corp., Japan) using a standardised setup. The subjects thenThe first system uses the five most discriminativeunderwent a diagnostic polysomnography (PSG) overnightfeatures from 71 craniofacial based on the SVM-REFtest.Following the study, PSG scoring was performed bytechnique [5]. These features are face width, mandibularexperienced accredited sleep technologists Sleep staginglength 2, binocular width, cranial base area (ax), andwas performed using standard definitions. Apnoea wascriocomental space distance.defined as complete airflow cessation for greater than 10The second and third systems use the selectedseconds with oxygen desaturation of at least 3% and/orcalibrated and uncalibrated features from the study byassociated with arousal. Hypopnoea was defined as aLee.et.al [3].reduction in amplitude of airflow or chest/abdominal wallLee’s calibrated system used a logistic regressionmovement greater than 50% of the baseline measurementmodel processing the following calibrated photographicfor more than 10 seconds with an accompanying oxygenmeasurements: face width, eye width, cervicomentaldesaturation of at least 3%, and/or associated withangle, and mandibular length. It obtained 76.1% ofarousals. AHI was calculated as the total number ofaccuracy, sensitivity of 86%, specificity of 59.1% andapnoeas and hypopnoeas per hour of sleep.area under curve of ROC of 0.82 for discriminanting3. Signal Processing sleep apnoea using an AHI threshold of 10. This systemresulted in the highest rate of true classification with theSeventy one features for each of the 180 photographs lowest number of variables [3].had previously been manually determined by an expert toLee’s uncalibrated system also used a logisticsupport the studies described in [2,3,4].regression model processing the following uncalibratedphotographic measurements: face width-eye width ratio,Our first step was to normalize the features into thecervicomental angle and mandibular-nasion angle 2. Thisinterval 0 to 1 by scaling the mean and the variance. Themodel achieved an accuracy of 71.1% with sensitivity offeatures were then ranked using a Support Vector Machine80.7%, specificity of 54.5% and area underROC curveReverse Elimination Feature (SVM-REF) algorithm [5]of 0.80 [3].and the optimal number of ranked features selected. TheThe measurements used in Lee’s and our system areresult of this algorithm was five selected features. Thesupport vector regression (SVR) algorithm was then used illustrated in Table 1 and Figure 1. Table 2. AHI estimation using craniofacial measures andclinical variables using RBF kernel.Feature set MAE CCLee’s uncalibrated features 13.3 0.54Lee’s calibrated features 13.5 0.535SVM-REFselected features 13.4 0.525 clinical features 14.6 0.42Table 3. Classification Results of OSA detection using radialbasis function kernel SVMFeature/ROCAccuracy Sensitivity SpecificityAUCPerformanceUncalibrated 68.3 87.7 34.8 0.75Calibrated 72.22 89.47 42.42 0.7773.3 90.4 43.9 0.78Selected71.67 89.47 40.91 0.78ClinicalNotes: Calibrated: Lee’s 4 calibrated features [3]; 5 Selected: SVM-REF 5Figure 2. Scatter plot of predicted AHI versus true AHI for 5selected features; Uncalibrated: Lee’s 3 uncalibrated features [3]; Clinical: 5 clinicalSVM-REF selected features.featuresA fourth system using clinical features was also considered. The clinical features were age, BMI, neck ? ? ?circumference, abdominal girth, and hip girth. {3.2.Support Vector RegressionSVM uses an-insensitive loss function to solve Using Lagrange multipliers in solving the dualregression problems. Support Vector Regression (SVR) problem leads to the followingfunction attempts to find a continuous function where training points lie within distance of the target values. The ? ? ? ?factor is chosen to balance the margin of error and generalisibility of the prediction function to unseen data [6, 7]. LibSVM has been used to calculate the regression ? measurement via support vectors [7]. The SVR modelsthe regression directly. Similar to support vector?classifiers [8], a key part of SVRs is producing ameasurement of similarity using a kernel function. The The approximation function isbasic concept of SVR is to discover a function which map the input training data closest to the targets obtained by solving (4) using in a way to obtain the most flat function [6,7]. A simple ?, linear functionis illustrated as follows, ? where the are the ideal outputs, is a learned The support vector regression is estimated as above inconstant and the weight vector is a linear combinationof training points. In order to obtain flatness, smaller (3) [6].weights,, are found through minimizing its norm,? ?, through a convex optimization problem,3.3.Parameter Optimisation ? ?LibSVM was used for training and testing of the proposed model [7]. A grid search was utilized for { optimizing the SVR and RBF parameters [9]. It employedEquation 2 is augmented by introducing a region fora 5-fold cross validation of comprehensive searching ofthe support vectors. This is created by using the negligentthe subset of hyperplanes and hyper-parameters tovariables,, with a constant cost function, optimize regression performance [9].which represents the cost between the flatness of function andthe deviation tolerance over,3.4.Performance Measures 5. Conclusion Two measures were used to evaluate the performance This paper has presented a novel way of estimating theof the model including Mean Absolute Error (MAE) of apnoea-hypopnoea index using craniofacial photographs.the estimated AHI measures and Pearson linear Five craniofacial measurements were selected from 71correlation coefficient (CC) between the predicted AHI manual craniofacial phenotype features determined fromand the expert determined AHI [7].frontal and profile photographs of a subject’s head. TheAccuracy, sensitivity, specificity, and area under the measurements were processed with a Support Vectorcurve for the receiver operator characteristic (AUC-ROC) Machine Regression algorithm to estimate AHI. Usingwere used to quantify classification performance.leave-one-record-out cross-validation the estimatedaccuracy of the algorithm was 73.3% with an area underROC of 0.78. The correlation coefficient of the estimated4. Results and Discussion AHI against the expert AHI was 0.52.The results of AHI estimation using craniofacialAcknowledgementsmeasure and clinical variables using the RBF kernel areshown in Table 2 for the three sets of craniofacialWe thank Prof. Peter Cistulli and Dr Kate Sutherlandphotograph features and the clinical features. A scatterof the University of Sydney for providing the database ofplot for the SVM-REF selected features is shown inphotographic measurements for this study.Figure 2. The classification results are shown in Table 3.The results in Table 2 show that the three craniofacialfeature sets achieved around 0.5 correlation with theReferencesexpert determined AHI. The scatter plot for the SVM- REF features shows while the AHI trend is correct there[1] Young T, Peppard P E, Gottlieb D J. Epidemiology ofobstructive sleep apnea: A population health perspective.is a high degree of variability on an individual determinedAmerican Journal of Respiratory and Critical Care MedicineAHI point. Encouragingly, the system had low level of2002;165:1217–1239false negatives (see Fig. 2) indicating that the system may[2] Lee R, Chan A, Grunstein R, Cistulli P. Craniofacialhave potential utility as a screening device. This resultphenotyping in obstructive sleep apnea-a novel quantitativewas also seen in the high level of specificity 90.4% inphotographic approach. Sleep 2009;32:37–45Table 3. The clinical features resulted in a lower[3] Lee R, Petocz P, Prvan T, Chan A, Grunstein R, Cistulli P.correlation of 0.42, demonstrating that the craniofacialPrediction of obstructive sleep apnea with craniofacialphotograph features had greater predicting power.photographic analysis. Sleep 2009;32:46–52Our classification results using a regression model are [4] Lee R, Sutherland K, Chan A, Zeng B, Grunstein R,comparable to Lee’s result of 76.1% [4] but with the Darendeliler A, Schwab R, Cistulli P. Relationship betweensurface facial dimensions and upper airway structures inadded ability of estimating the severity of sleep apnoea. obstructive sleep apnea. Sleep 2010;33:1249–1254While our system has a way to go before it is a[5] Guyon I. Gene Selection for Cancer Classification. Machineclinically useful system, we’ve shown that a highlyLearning 2002;46:389–422 convenient, low cost, day-time test can be used to predict[6]SchölkopfB. A tutorial on support vector regression. Statthe severity of sleep apnoea. With further development,Comput 2004;14:199–222 our system has potential as primary diagnostic tool to[7] Chang C, Lin C. LIBSVM?: A Library for Support Vectortackle the societal burden of undiagnosed sleep apnoea.Machines. ACM Trans Intell Syst Technol 2011;2:1–39Future work will look at restricting the feature pool to[8] Vapnik V. Statistical Learning Theory. New York: Wileyand Sons, 1998:1–740the set of measurements not requiring calibration (e.g.[9] Espinoza-cuadros F, Fernández-pozo R, Toledano DT,angles and relative distances) and choosing an optimalAlcázar-ramírez JD, López-gonzalo E, Hernández-gómezsubset. This would remove the need to use the calibrationLA. Speech Signal and Facial Image Processing forwashers used on the forehead and the cheek shown in Fig.Obstructive Sleep Apnea Assessment. Comput Math1. Our selected landmarks can also serve as a baseline for Methods Med 2015;2015:1–13 a fully automatic photographic analysis system for OSAdetection, where the key craniofacial indicators are to beAddress of correspondence: Philip de Chazalidentified automatically using image processingSchool of Electrical and Information Engineering, Building J03, The University of Sydney,NSW, 2006, Australia.algorithms. This could create the opportunity of using [email protected], ubiquitous technology such as smart camera phonesto perform the apnoea screening. "