Detecting Low Frequency Oscillations in Cardiovascular Signals Using Gradient Frequency Neural Networks

Thomas Kaplan1 and Elaine Chew2
1Queen Mary University of London, 2CNRS-STMS (IRCAM)


Gradient Frequency Neural Networks (GFNNs) have been applied successfully to detect pulse and meter (hierarchical groupings of pulses) in complex music audio signals having polyrhythms and syncopation. Here, we apply GFNNs to the detection of low frequency oscillations in several cardiovascular signals, with the aim to capture Mayer waves---low frequency oscillations in arterial pressure that can be linked to the baroreceptor reflex. The cardiovascular time series is treated as music audio for analysis; the ECG signal is processed as a WAV file, and R-R intervals converted to a MIDI file. GFNNs are neural oscillator networks that process a time-varying signal, where each network layer consists of many nonlinear oscillators with different natural frequencies along some gradient. Its nonlinearity offers the advantage of high sensitivity at low stimulus amplitudes, compared to linear amplitude responses for weak signals. In accordance with previous observations, GFNNs entrained with pronounced components around 0.025-0.35Hz to R-R intervals of meditators from the PhysioBank Exaggerated Heart Rate Oscillations database (comprising of recordings of 8 Chi meditators, 5 women and 3 men, and 4 Kundalini Yoga meditators, 2 women and 2 men). When applied to an 18-hour Holter recording, GFNN entrainment was observed around various documented frequencies of physiological interest---very low frequency (~0.04Hz), low frequency (~0.1Hz), and high frequency (~0.25Hz). These results are compared to traditional time-domain and frequency-domain techniques, and the effect of varying the GFNN parameter for system nonlinearity is demonstrated. GFNNs present a novel approach to the detection and study of cardiovascular oscillations, inspired by auditory rhythm perception.