Representing classifier confidence in the safety critical domain: an illustration from mortality prediction in trauma cases

Abstract

This work proposes a novel approach to assessing confidence measures for software classification systems in demanding applications such as those in the safety critical domain. Our focus is the Bayesian framework for developing a model-averaged probabilistic classifier implemented using Markov chain Monte Carlo (MCMC) and where appropriate its reversible jump variant (RJ-MCMC). Within this context we suggest a new technique, building on the reject region idea, to identify areas in feature space that are associated with “unsure” classification predictions. We term such areas “uncertainty envelopes” and they are defined in terms of the full characteristics of the posterior predictive density in different regions of the feature space. We argue this is more informative than use of a traditional reject region which considers only point estimates of predictive probabilities. Results from the method we propose are illustrated on synthetic data and also usefully applied to real life safety critical systems involving medical trauma data.