Evaluating Frequency-Domain Features for CNN-based Sleep Apnea Detection Using Respiratory Effort Signals

Abstract

Sleep apnea syndrome is a sleep-related breathing disorder characterized by repeated episodes of apnea or hypopnea. In recent years, deep learning models have been increasingly applied to detect such abnormalities from respiratory motion signals. Most existing approaches process these signals directly in the time domain, with relatively limited attention given to frequency-domain features that reflect the periodicity of respiratory patterns. In this study, we investigate the effectiveness of frequency-domain features—derived from Fourier and wavelet transforms of thoracic and abdominal respiratory signals— for detecting respiratory abnormalities. We conducted experiments using polysomnographic data from 100 subjects in the SHHS2 dataset and developed several CNN-based models that utilize time-domain, frequency-domain, or combined features as input. The results showed that integrating time- and frequency-domain features improved detection performance compared to using frequency-domain features alone. However, further analysis revealed that key discriminative information for identifying abnormal respiratory events is more closely related to transient and localized changes in motion patterns than to variations in periodicity. These findings suggest that while frequency-domain features provide supplementary cues, time-domain representations remain critical for the accurate detection of respiratory abnormalities.