Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
517429 | Journal of Biomedical Informatics | 2010 | 9 Pages |
An important problem in the Intensive Care is how to predict on a given day of stay the eventual hospital mortality for a specific patient. A recent approach to solve this problem suggested the use of frequent temporal sequences (FTSs) as predictors. Methods following this approach were evaluated in the past by inducing a model from a training set and validating the prognostic performance on an independent test set. Although this evaluative approach addresses the validity of the specific models induced in an experiment, it falls short of evaluating the inductive method itself. To achieve this, one must account for the inherent sources of variation in the experimental design. The main aim of this work is to demonstrate a procedure based on bootstrapping, specifically the .632 bootstrap procedure, for evaluating inductive methods that discover patterns, such as FTSs. A second aim is to apply this approach to find out whether a recently suggested inductive method that discovers FTSs of organ functioning status is superior over a traditional method that does not use temporal sequences when compared on each successive day of stay at the Intensive Care Unit. The use of bootstrapping with logistic regression using pre-specified covariates is known in the statistical literature. Using inductive methods of prognostic models based on temporal sequence discovery within the bootstrap procedure is however novel at least in predictive models in the Intensive Care. Our results of applying the bootstrap-based evaluative procedure demonstrate the superiority of the FTS-based inductive method over the traditional method in terms of discrimination as well as accuracy. In addition we illustrate the insights gained by the analyst into the discovered FTSs from the bootstrap samples.