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

Area of Science:

• Medical Imaging and Machine Learning
• Sleep Medicine and Diagnostics
• Biomedical Engineering

Background:

• Obstructive sleep apnea (OSA) is significantly underdiagnosed due to limited access to polysomnography (PSG).
• Current diagnostic methods for OSA are complex and resource-intensive.
• There is a need for more accessible and efficient OSA screening tools.

Purpose of the Study:

• To evaluate the performance of machine learning (ML) models in predicting OSA using 3D maxillofacial shapes.
• To compare the diagnostic accuracy of ML-based 3D shape analysis with established questionnaires (BERLIN and NoSAS).
• To explore the potential of 3D geometric morphometrics as a rapid screening tool for OSA.

Main Methods:

• 3D maxillofacial scans were acquired from 267 Caucasian men with suspected OSA.
• Participants underwent polysomnography (PSG) for OSA diagnosis.
• 13 supervised ML algorithms were trained and tested on processed 3D craniofacial scans; results were compared to BERLIN and NoSAS questionnaire performance.

Main Results:

• ML analysis of 3D craniofacial shapes achieved 56% specificity for OSA (AHI≥15), outperforming the BERLIN (50%) and NoSAS (40%) questionnaires.
• ML models demonstrated 80% sensitivity for OSA detection, compared to NoSAS (approx. 90%) and BERLIN (61%).
• Combining 3D geometric morphometrics with anthropometric data improved the area under the receiver operating characteristic curve (auROC) to 0.75.

Conclusions:

• 3D geometric morphometrics combined with ML offers a promising approach for OSA screening.
• This method presents a rapid, efficient, and potentially inexpensive alternative for OSA detection.
• Further development could enhance accessibility to OSA diagnosis.