| Hemodynamic indices that can be estimated from blood pump flow were studied for the detection of ventricular suction. Results demonstrate that an individual index alone will not provide a reliable suction detector, but combination of indices can improve the performance for detection. Combination of indices based on fuzzy, Dempster-Shafer and Bayesian theory were developed. The Bayesian model provided the best overall performance, but the fuzzy and Dempster-Shafer models, which incorporate uncertainty, can reduce the number of explicit errors by creating uncertain classifications.; A new detection system, the certainty-weighted detection system (CWDS), is described. The CWDS contains two components, one that provides a binary crisp decision and one that provides a continuous measure of uncertainty. The measure of uncertainty is high when physiologic indices of suction are ambiguous, and it can provide uncertain classifications by means of an additional decision rule. Unlike fuzzy and Dempster-Shafer systems, the CWDS is based directly on probability density functions. The binary crisp decision component of the CWDS provides a model whose performance can change from being similar to a distributed Bayesian detector to being similar to the classical Bayesian detector by changing the definition of uncertainty. The CWDS with uncertain classifications was compared to fuzzy and Dempster-Shafer approaches, using a definition of cost that differentiates between explicit errors and uncertain classifications. The use of uncertain classifications in the CWDS can reduce the cost to less than that of the CWDS without uncertainty, if the cost of errors in uncertain classifications is less than the cost of error in crisp classifications. Simulation data and in vivo data were used to demonstrate this performance. |
