Time-Frequency Machine Learning Transfer Function for Central Pressure Waveforms

Abstract Background Clinical studies show that pulsatile hemodynamics and pressure waveform analysis are valuable for the diagnosis and prognosis of hypertension and heart failure. While generalized transfer functions (GTF) have shown clinical significance, some studies report limitations with GTF in capturing central pulsatile hemodynamics. This study introduces a hybrid time-frequency, machine learning-based transfer function that reconstructs central pressure waveforms from peripheral measurements, accurately capturing central pulsatile hemodynamics and arterial wave-based information. Methods Our method uses Fourier harmonics for approximating the pressure waveform. The model is trained on these harmonics using a feed-forward neural network (FNN) with a custom time-domain cost-function that captures full temporal dynamics of physiological events during a cardiac cycle. The final hybridized-FNN transfer function model is trained, tested, and validated on data from the Framingham Heart Study (6,698 participants). Results Our method produces carotid waveforms with median normalized mean squared error (%NMSE) values of 0.09 and 0.10 for brachial and radial inputs, compared to 0.42 and 0.26 for GTF, with similar accuracy improvements in other metrics. Correlation coefficients for the first and second forward wave times and amplitudes are 0.97, 0.93, 0.82, and 0.79 with brachial input, and 0.97, 0.92, 0.87, and 0.80 with radial input, versus as low as 0.22 and 0.31 for GTF. Overall, our method significantly improved correlations across similarity, morphology, and wave-based parameters. Conclusions Our Hybridized FNN transfer function approach enables robust calculation of the central arterial pressure waveform from a single measured peripheral waveform, preserving key physiological sequences in a cardiac cycle.