31 Towards the identification of independent measures of heart rate variability Andrea Bravi , Christophe Herry , Daphne Townsend , Geoffrey Green , Tim Ramsay , Andre Longtin , Andrew Seely a,b,c University of Ottawa, Ottawa, Ontario, Canada Ottawa Hospital Research Institute, Ottawa, Ontario, Canada Therapeutic Monitoring Systems (TMS), Inc., Ottawa, Ontario, Canada Objectives:Aplethora ofmeasures is applied to the study of heart rate variability (HRV). The physiological explanation of those measures remains unclear; most measures track similar phenomena, and high correlation exists among them. Eliminating redundancy for patient classification is a principal objective (eg, distinguishing patient who will develop infection from those who will not). Indeed, redundant featuresmay impair a classifier derived fromHRV, reducing its ability to produce reliable decision boundaries. This work addresses this issue, proposing a robust procedure to identify a set of uncorrelated measures of arbitrary size (ie, tuned for a specific application). Methods: This study uses a data set tracking the diurnal variations in HRV in 42 ambulatory healthy subjects with heterogeneous age and fitness levels (ie, body mass index and clinical history). For each subject, 93 measures of HRV were tracked over time through a windowed analysis (5-minute window size, no overlap for 24 hours) of the relative R-R interval time series. For each pair of measures, the amount of shared information was estimated by extracting their Spearman correlations. Then, through bootstrapping (ie, iterative recalculation), the population statistics were estimated, obtaining the mean and SE (andmedian and interquartile range error) of the values of nonlinear correlation between each pair of measures. A greedy optimization algorithm was then used to create a decision surface showing the tradeoff between the number of measures to select and their correlation; this enables the determination of an optimal set of uncorrelated variability measures. Results: The population estimates of correlation showed an error of ±0.1 (with 95% confidence interval), therefore providing a robust estimate of the relationship between themeasures. From the decision surface created by the optimization algorithm for this data set, we selected to keep 4 measures of variability (this number could be appropriate for a study trying to classify 2 populations of 50 patients each). The maximum correlation among the group of measures was 0.01 (ie, none of the pairwise correlations were higher than that value). This resulted in the selection of (1) average of the differential of the R-R interval, (2) kurtosis, (3) forbidden words of the symbolic dynamics, and (4) X-intercept of the power law. Conclusions: The proposed method represents a potential means to select a priori a set of uncorrelated measures of variability. Given their lack of redundancy, this set represents an ideal starting set to solve a patient classification task. The validation of the approach on independent data sets is required. http://dx.doi.org/10.1016/j.jcrc.2012.10.047