Sleep Apnea Detection Directly from Unprocessed ECG through Singular Spectrum Decomposition

ECG-based detection of sleep apnea is generally based on heart rate related indices. Computation of these indices requires an ECG record to be pre-processed and the R-peak locations to be estimated. This study proposes a novel method to detect minute-by-minute sleep apnea episodes directly from an unprocessed ECG through singular spectrum decomposition (SSD). Given an ECG record, SSD was applied to non-overlapping segments and the dominant frequency (DF) of the component in the frequency range 0.02-0.5 Hz was estimated. Each segment was binary classified based on the corresponding DF (1, if DF larger than a defined threshold, 0 otherwise). For every minute, the sum of the corresponding binary values was then used to classify that minute as normal or apnea. Validation was based on the learning set of the Apnea-ECG Database. Two segment lengths, 10s and 20s, were tested, and K-fold cross-validation was used to determine the optimal values for threshold and sum, and for performance analysis. The 20s segment-based analysis proved to be more reliable and provided a sensitivity and a specificity of 67% and 52%, respectively. Although performance of the proposed model is still unsatisfactory, the preliminary results reported in this study suggest that detection of sleep apnea directly from unprocessed ECG may be possible.