Empirical Mode Decomposition of simulated and real ultrasonic Doppler signals of periodic fetal activity

Abstract

Simulated signals comprising components (trains of Gaussian packets) resulting from cardiac movements and from pseudorespiratory movements with added white noise were submitted to Empirical Mode Decomposition. The increase of sampling frequency fs (from 0.5kHz to 5kHz) for given signal to noise ratio SNR moves signal components toward higher order intrinsic mode functions (IMFs) and increases their number. The increase of the SNR (from −5dB to 10dB, fixed fs) moves the signal components to lower order IMFs. The separation of components is most efficient for SNR≥5dB and fs not exceeding 1kHz, for lower SNRs fs should be at least 2kHz. SNR=∞ results in erroneous decomposition and therefore limited noise level is beneficial. Recommended number of sifting iterations is 10. Fetal data obtained using 2MHz emission frequency and sampled at 2kHz were decomposed. The cardiac signal always appears in IMF3, frequently also in IMF1 and IMF2. The pseudobreathing signal, appearing mainly in IMF4-6, is easy to separate. Signals resulting from fetal displacements due to maternal respiration appear in IMF7 or IMF8. The EMD performs better than the classic linear filtering as a tool for separation of the pseudorespiration signals and provides inferior results in terms of separation of the cardiac signals.