Predict-OSA: Integrative Multimodal-based Early Prediction of Sleep
Apnea using Single-lead ECG Signal
Abstract
Obstructive sleep apnea (OSA) is a sleep condition characterized
by the partial or full interruption of breathing while a person is
asleep. Several techniques have been suggested for the automated
recognition of OSA events. However, there has been little research on
the prediction of these events using prior data, which is valuable for
the advancement of medical devices that observe respiration while
sleeping. In this research, we have proposed four techniques for
forecasting OSA events by using a fusion of time-frequency
representation (TFR) of electrocardiogram (ECG) signals and deep neural
networks (Gabor-CNN, Wavelet-CNN, Wigner-CNN, and STFT-CNN). Overall,
these techniques are called Predict-OSA. These techniques use raw
single-lead ECG data collected at a sampling rate of 100 Hz without the
inclusion of any manually extracted features. By analyzing the preceding
120 seconds of data, we can predict the occurrences of OSA (apnea) and
regular breathing episodes in a lead time of 60 seconds. The findings
obtained from analyzing a dataset consisting of more than 34270
60-second segments from 70 ECG recordings indicate that all four models
had a high level of accuracy, surpassing 83%. The Wigner-CNN model
demonstrated superior performance, achieving accuracy, specificity, and
sensitivity of 89.75%, 91.89%, and 86.25%, respectively. The findings
indicate that OSA occurrences may be reliably anticipated by analyzing
single-lead ECG signals, which creates possibilities for designing
devices that can prevent the incidence of OSA events from occurring
during sleep. Our algorithms may be implemented in devices with limited
storage space due to their ability to handle low sampling rates. This
makes them well-suited for use in at-home environments.