A Multi-Feature Fusion Framework for Automated Classification of Obstructive and Central Hypopneas in Polysomnography

This study aims to develop and validate a fully automated, interpretable machine learning system for the precise classification of obstructive (OH) and central (CH) hypopneas in polysomnography (PSG), directly addressing the critical issue of subjectivity and inconsistency in manual visual scoring.

We computationally operationalized the hierarchical decision logic of clinical standards (AASM & Randerath) by extracting a comprehensive panel of 15+ multi-modal features from standard PSG signals. These features quantitatively capture key physiological phenomena: inspiratory flow limitation (IFL), thoraco-abdominal paradox, event termination pattern, arousal timing, and sleep-stage association. An interpretable confidence-guided hierarchical classification (CGHC) framework was proposed for binary hypopnea classification.

The model was evaluated on a dataset of 2,509 hypopnea events. It achieved an accuracy of 85.29% and a Cohen's kappa of 0.71, significantly outperforming manual scoring consistency and demonstrating superior performance against a consensus expert ground truth. Feature importance analysis confirmed IFL as the primary predictor, validating the clinical decision hierarchy.

The results demonstrate that our automated system can accurately and reliably replicate expert clinical judgment for hypopnea subtyping.

This tool provides the first robust and interpretable solution for automating complex hypopnea classification, promising to enhance diagnostic precision, standardize scoring, and facilitate personalized therapy for sleep-disordered breathing.