This paper presents an empirical comparison of two temporal abstraction procedures, that were applied to derive predictive features for a prediction problem in intensive care medicine. The first procedure employs knowledge from practitioners to derive qualitative patterns of state changes; the second procedure searches through a large number of data summaries to discover those that have predictive value. The derived features were used to predict whether postsurgical patients would need mechanical ventilation longer then 24h. The data-driven temporal abstraction procedure was found to provide more informative predictors, resulting in better predictions.