Temporal Classification of Medical Events

We investigate the task of assigning medical events in clinical narratives to discrete time-bins. The time-bins are defined to capture when a medical event occurs relative to the hospital admission date in each clinical narrative. We model the problem as a sequence tagging task using Conditional Random Fields. We extract a combination of lexical, section-based and temporal features from medical events in each clinical narrative. The sequence tagging system outperforms a system that does not utilize any sequence information modeled using a Maximum Entropy classifier. We present results with both hand-tagged as well as automatically extracted features. We observe over 8% improvement in overall tagging accuracy with the inclusion of sequence information.