Machine learning and geometric morphometrics to predict obstructive sleep apnoea from 3D craniofacial scans

Abstract

<b>Rationale:</b> Obstructive sleep apnoea (OSA) remains massively underdiagnosed, due to reduced access to polysomnography (PSG), the highly complex gold standard for diagnosis. Performance scores in predicting OSA are evaluated for machine learning analysis applied to 3D maxillofacial shapes. <b>Methods:</b> The 3D maxillofacial shapes were scanned on 280 Caucasian men with suspected OSA. All participants underwent single night in-home or in-laboratory sleep testing with PSG (Nox A1, Resmed, Australia), with concomitant 3D scanning (3D scanner Sense v2, 3D systems corporation, USA). Anthropometric data, comorbidities, medication, BERLIN, and NoSAS questionnaires were also collected at baseline. The PSG recordings were manually scored at the reference sleep centre. The 3D craniofacial scans were processed by geometric morphometrics, associated with supervised machine learning (ML) algorithms. Results for OSA recognition by ML models were compared to or better than those obtained by traditional BERLIN and NoSAS questionnaires, in terms of area under the Receiver Operating Characteristic curve (auROC). <b>Results:</b> All valid scans (n=267) were included in the analysis (patient mean age: 59±9 years; BMI: 27.4 kg/m²). For PSG-derived AHI≥15 events/h, the 56% specificity obtained for ML analysis of 3D craniofacial shapes was higher than for both of the questionnaires (Berlin: 50%; NoSAS: 40%). The auROC score was further improved when 3D geometric morphometrics were combined with patient anthropometrics (auROC=0.75). <b>Conclusion:</b> The combination of 3D geometric morphometrics with machine learning is therefore proposed as a rapid, efficient, and inexpensive screening tool for OSA.