Respiratory impedance spectral estimation for digitally created random noise

Measurement of respiratory input mechanical impedance (Zrs) is noninvasive, requires minimal subject cooperation, and contains information related to mechanical lung function. A common approach to measure Zrs is to apply random noise pressure signals at the airway opening, measure the resulting flow variations, and then estimate Zrs using Fast-Fourier Transform (FFT) techniques. The goal of this study was to quantify how several signal processing issues affect the quality of a Zrs spectral estimate when the input pressure sequence is created digitally. Random noise driven pressure and flow time domain data were simulated for three models, which permitted predictions of Zrs characteristics previously reported from 0–4, 4–32, and 4–200 Hz. Then, the quality of the Zrs estimate was evaluated as a function of the number of runs ensemble averaged, windowing, flow signal-to-noise ratio (SNR), and pressure spectral magnitude shape |P(jω)|. For a |P(jω)| with uniform power distribution and a SNR<100, the 0–4 Hz and 4–200 Hz Zrs estimates for 10 runs were poor (minimum coherence γ2<0.75) particularly where Zrs is high. When the SNR>200 and 10 runs were averaged, the minimum γ2 >0.95. However, when |P(jω)| was matched to |Zrs|, γ2 > 0.91 even for 5 runs and a SNR of 20. For data created digitally with equally spaced spectral content, the rectangular window was superior to the Hanning. Finally, coherence alone may not be a reliable measure of Zrs quality because coherence is only an estimate itself. We conclude that an accurate estimate of Zrs is best obtained by matching |P(jω)| to |Zin| (subject and speaker) and using rectangular windowing.