Session S53.2

Detection Of Sleep Apnea In Single Channel ECGs From The PhysioNet Data Base

M. Schrader, C. Zywietz, V. von Einem, B. Widiger,
G. Joseph

Medical School Hannover
Hannover, Germany

The work reported here refers to the CiC Challenge 2000 on detecting sleep apnea using single lead ECGs.
1. The ECG data were downloaded from PhysioNet including the QRS and apnea annotation data. 2. Since the ECGs provided had a sampling rate of 100 samples/s we interpolated the ECGs by means of a 10s second order spline polynome to a sampling rate of 500 samples/s in order to make it possible to use our own QRS detector and for a better QRS-peak identification. 3. All data were then processed by our arrhythmia (Holter) ECG program after its modification for single channel analysis. 4. QRS detection was compared between the PhysioNet annotation and our results. QRS localizations were corrected by visual inspection. 5. After that data sections with premature ventricular and supraventricular beats were marked and excluded from further analysis. 6. Mean heart rates and heart rate variability (HRV) were computed for intervals of 10s, 30s, 60s, 3 min, 5 min ... and for overall. 7. For a "walk through" search Fourier and Discrete Harmonic Wavelet analysis for LF (0.04 - 0.15 Hz); HF (0.15 - 0.4 Hz) and brady-tachycardia arousal (0.02 - 0.075 Hz) bands is performed. 8. From 10s record sections ECG mean cycle measurements were extracted. 9. Using the 35 annotated records as learning set a multivariate classification procedure was designed. Application of this procedure to the learning set revealed a sensitivity of 85% for detection of sleep apnea ECGs, specificity was 73%. 10. Application of this procedure to the test set gave apnea episodes in 18 ECGs. Nine ECGs were classified as non-apnea ECGs and for eight ECGs no classification was given.
We believe that the analysis of single channel ECGs provides only limited information because QRS and ST-T changes can not be associated with respiration movement as it is possible in, for example, three orthogonal leads. Nevertheless, we hope to improve the performance of our algorithm by improved analysis of frequency components in HRV by means of better adapted interval window settings.