Dynamic selection of models for a ventilator-management advisor

A ventilator-management advisor (VMA) is a computer program that monitors patients who are treated with a mechanical ventilator. A VMA implements a patient-specific physiologic model to interpret patient data and to predict the effects of alternative control settings for the ventilator. Because a VMA evaluates its physiologic model repeatedly during each cycle of data interpretation, highly complex models may require more computation time than is available in this time-critical application. On the other hand, less complex models may be inaccurate if they are unable to represent a patient's physiologic abnormalities. For each patient, a VMA should select a model that balances the tradeoff of prediction accuracy and computation-time complexity. I present a method to select models that are at an appropriate level of detail for time-constrained decision tasks. The method is based on a local search in a graph of models (GoM) for a model that maximizes the tradeoff of computation-time complexity and prediction accuracy. For each model under consideration, a belief network computes a probability of model adequacy given the qualitative prior information, and the goodness of fit of the model to the data provides a measure of the conditional probability of adequacy given the quantitative observations. I apply this method to the problem of model selection for a VMA. I describe an implementation of a graph of physiologic models that range in complexity from VentPlan, a simple model with 3 compartments, to VentSim, a multicompartment model with detailed airway, circulation and mechanical ventilator components.(ABSTRACT TRUNCATED AT 250 WORDS)