CTC-Based Apnea Hypopnea Index Estimation using Single-Channel ECG

Abstract

This paper presents the first electrocardiogram-only system that estimates the Apnea-Hypopnea Index (AHI) using Connectionist Temporal Classification (CTC) loss. CTC trains the network from the nightly count of apnea events, eliminating the frame-level time stamps demanded by conventional crossentropy (CE) approaches and sharply reducing annotation effort. A ContextNet spectrogram encoder followed by a Transformer is trained with either CTC or CE; our CTC model surpasses the performance of the CE baseline, showing that alignment-free supervision can in fact enhance model robustness and accuracy. Because ECG reactions lag the actual airway obstruction by several seconds, CTC's built-in timing flexibility is especially advantageous for accurately modeling this delayed physiological response. The proposed method therefore enables accurate, annotation-efficient, and wearable-friendly screening for sleep apnea.