Machine Learning-Based Screening for Obstructive Sleep Apnea Using Photographic Craniofacial and Speech-Derived Acoustic Features

Obstructive sleep apnea (OSA) is a prevalent disorder for which accessible non-invasive screening approaches remain of interest. This study evaluated photographic craniofacial descriptors and speech-derived acoustic features for Apnea-Hypopnea Index (AHI) estimation and binary OSA screening. Data from 109 adults were analyzed using unsegmented and temporally segmented speech representations. Seven craniofacial, acoustic, and combined feature configurations were evaluated using nested cross-validation. For regression, the lowest MAE was 14.242 events/h using segmented continuous-speech features with Random Forest. For classification, the highest AUC-ROC was 0.766 using unsegmented craniofacial features with a multilayer perceptron classifier, while the best segmented result was 0.756 using craniofacial plus continuous-speech features. Craniofacial-only features in the segmented representation achieved a similar AUC-ROC of 0.753, indicating no clear discriminative advantage from adding acoustic features. Several classifiers showed high sensitivity but comparatively low specificity. Sensitivity analysis excluding the four participants with the highest AHI values showed similar classification patterns, although regression performance was more sensitive to cohort composition. Overall, craniofacial descriptors provided the most consistent predictive contribution, whereas acoustic features showed model- and representation-dependent value for OSA screening.