- The increasing availability of temporal data from Electronic Health Records (EHR) provides exceptional opportunities for the prediction of clinical outcome events. The EHR aggregates data from many sources into a centralized database. However, the nature of EHR data is sparse, non-randomly missing and heterogeneous, which is very challenging to analyze. We propose the use of temporal abstraction to transform the data into symbolic time intervals series. Then we use KarmaLego, a fast time intervals mining algorithm, to discover frequent Time Intervals Related Patterns (TIRPs) that are used as features to predict the outcome event. In this study we rigorously evaluate various aspects in the KarmaLego Outcome Events Prediction framework on an extraction of 32,168 patients from Columbia University Medical Center focusing on 6 procedures as outcome events. Our results show that the use of TIRPs for prediction significantly outperforms using only static concepts, and the use of our two TIRPs metrics for features representation outperform the use of the default Binary TIRP representation.