Data transforms for spectral analyses of heart rate variability.

Autoregressive and fast Fourier transform spectral analyses of high-frequency heart rate variability (HF-HRV) result in exponentially-distributed values that make standard parametric statistical analyses problematic. In this paper, we evaluate three transforms of raw HF-HRV spectral power. Two occur commonly in the literature (a natural log [ln] transform and a reactivity transform); a third is novel (a "percent deviation from the mean" transform). A single data set was used, with each subject providing two data points and for which we predicted a significant difference in HF-HRV power. We quantified the effect size of each transform by noting the percentage of (non)overlap between the +/- 1 standard errors surrounding the two period means, with less overlap indicating a stronger effect. Overlap was 19.2% in the raw data (Fig 1b.), 3.7% in the ln transform (Fig. 2b), -57.1% in the reactivity transform (Fig. 3b), and -70.2% in the percent deviation transform (Fig. 4b). The percent deviation transform resulted in more normally-distributed data than the reactivity transform and more tightly distributed data than the ln transform, making it a favorable choice for investigators.