RR Interval Time Series Modeling: The PhysioNet/Computers In Cardiology Challenge 2002
Massachusetts Institute of Technology
Cambridge, MA, USA
Simulating a realistic sequence of RR intervals is a task made difficult by the intricate interdependencies of interval fluctuations at scales ranging from seconds to many hours. Models that can generate such simulations may be useful not only for providing time series with known properties for evaluation of novel analytic methods, but also for providing insight into mechanisms underlying heart rate variability. To stimulate the creation and exchange of high-quality models of inter-beat interval variability, PhysioNet and Computers in Cardiology have sponsored an open on-line competition devoted to this topic, the latest in a series of annual challenges for the research community.
Participants in this year's PhysioNet/CinC Challenge were invited to create and evaluate software to generate RR interval time series. In the first event of the Challenge, seven participants entered one or more generators, and we provided several additional generators. We compiled each generator and used it to generate two series with lengths randomly chosen between 20 and 24 hours, using different random seeds for initialization in each case. To these, we added a roughly equal number of real RR interval time series with a similar distribution of lengths, to obtain a dataset of 50 series. In the six days following the posting of the challenge dataset, seventeen participants in the second event (including all seven who had entered the first event) attempted to identify which of the 50 series were synthetic. Most were quite successful, including six who were able to classify all of the series without errors, demonstrating the difficulty of producing highly realistic models of RR variability. The classifications obtained in the second event determined the rankings of the generators submitted in the first event.
The C-language sources for all of the submitted generators will be made freely available on PhysioNet following the conference. This collection of models will be an important resource for future investigations that require synthetic RR or heart rate time series, and for development of even more realistic generators.