This subproject is one of many research subprojects utilizing theresources provided by a Center grant funded by NIH/NCRR. The subproject andinvestigator (PI) may have received primary funding from another NIH source,and thus could be represented in other CRISP entries. The institution listed isfor the Center, which is not necessarily the institution for the investigator.Our recent discovery of a close connection between the clinical concept of differential diagnosis and the multimodality in posterior distributions arising from assimilating physiological observations using a mechanistic mathematical model of the cardiovascular system (From Inverse Problems in Mathematical Physiology to Quantitative Differential Diagnoses, Zenker et al, PloS Computational Biology 2007, in press) suggests that this approach may yield a useful quantitative decision aid in data-rich fields of medicine such as critical care. We are in the process of implementing this computationally expensive approach with both animal experimental data and clinical data. The purpose of this exploratory proposal is to prepare for a future MRAC/LRAC application. We will pursue 2 specific aims: 1) determine which Teragrid ressources are best suited for our coarse grained computational tasks (performing Markov Chain Monte Carlo (MCMC) sampling on thousands of individual datasets for relatively low dimensional parameter spaces), and 2) perform testing and scalability analysis for a novel parallelized MCMC algorithm tailored to true high performance computing environments. In all cases, the underlying models for which state/parameter space inference is performed consist of highly nonlinear systems of ordinary differential equations describing the cardiovascular system with 19 to hundreds of parameters /initial conditions. The models are implemented in C and integrated using CVODE(S) from the SUNDIALS package. The datasets to be assimilated range from extremely sparse data for large numbers of experiments (6 scalar values for approx. 10000 experiments) to high resolution timeseries data for a smaller number of experiments (days of 250 Hz multi channel data for a few dozen individuals).
Under specific aim 1, a standard Metropolis Hastings algorithm, as well as coarse grained parallel versions of varieties with improved convergence properties (Parallel Tempering, Normal Kernel Coupler) that we have developed in our local PittGrid distributed computing environment will be tested on various TeraGrid cluster ressources.
Under specific aim 2, we will perform scalability testing of a novel, extremely parallel, auto-adaptive variant of Metropolis-Hastings with significant interprocess communication requirements currently under development in our group. This algorithm, which will be of generic applicability for high dimensional sampling problems, will be implemented in Charm++ to exploit dynamic load balancing capabilities and subjected to scalability testing on MPP resources (BigBen, Ranger if available).
Showing the most recent 10 out of 292 publications