Apnoea-hypopnoea index estimation using craniofacial photographic measurements

This paper presents a novel way of estimating the apnoea-hypopnoea index (AHI) using craniofacial photographs. We compared the correlation and classification performance of the photograph-determined AHI against expert-determined AHI for a number of selected measurement sets. Our best performing system used five craniofacial measurements selected from 71 manual craniofacial phenotype features, which had been determined from frontal and profile photographs of a patient's head and neck. The measurements were processed with a Support Vector Machine Regression algorithm to estimate AHI. The best features included face width, mandibular length, binocular width, cranial base area, and criocomental space distance. A database of 114 subjects with OSA (AHI>10/h) and 66 controls (AHI <;10/h) was used for algorithm development and testing. Leave-one-record-out cross-validation was used to estimate performance. The Pearson correlation was 0.52 for the AHI estimation. Classification performed using an AHI threshold of 10 events per hour, resulted in an estimated accuracy of the algorithm of 73.3% with an area under the ROC of 0.78.