Three-class obstructive sleep apnea severity assessment: a parallel AHI and ODI explainable artificial intelligence...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Machine learning for Obstructive Sleep Apnea (OSA) diagnosis faces limitations: generic features, sole reliance on Apnea-Hypopnea Index (AHI), poor multi-class grading, and lack of mechanistic insight.
• This study addresses these gaps by incorporating craniofacial and intraoral metrics into a predictive framework.

Purpose of the Study:

• To develop an interpretable machine learning (ML) framework for three-class OSA severity classification.
• To evaluate the framework's performance using both AHI and Oxygen Desaturation Index (ODI).
• To assess the added value of anatomical predictors in OSA severity classification.

Main Methods:

• A prospective, single-center study of 233 treatment-naïve adults using in-laboratory polysomnography (PSG).
• Predictor variables included demographic, anthropometric, questionnaire-based, and novel craniofacial/intraoral metrics.
• An Artificial Neural Network (ANN) was trained for three-class severity classification, with interpretability assessed using SHAP analysis.

Main Results:

• The AHI-based ANN model achieved 87.2% overall accuracy; the ODI-based model achieved 76.6% accuracy.
• SHAP analysis identified craniofacial features (e.g., V-shaped maxillary arch) and clinical factors (e.g., Mallampati score) as influential predictors.
• The anatomy-inclusive ANN significantly outperformed an ablation model (87.2% vs. 72.3% accuracy).

Conclusions:

• An interpretable ML framework integrating anatomical data can effectively classify OSA severity and provide clinical insights.
• The framework moves beyond AHI-only prediction, highlighting hypoxic burden as a key target.
• Further validation in diverse populations is necessary to confirm clinical utility.