Annotating and Detecting Medical Events in Clinical Notes

Annotating and Detecting Medical Events in Clinical Notes Prescott Klassen, Fei Xia, and Meliha Yetisgen University of Washington PO Box 352425, Seattle, WA 98195, USA {klassp, fxia, melihay}@uw.edu Abstract Early detection and treatment of diseases that onset after a patient is admitted to a hospital, such as pneumonia, is critical to improving and reducing costs in healthcare. Previous studies (Tepper et al., 2013) showed that change-of-state events in clinical notes could be important cues for phenotype detection. In this paper, we extend the annotation schema proposed in (Klassen et al., 2014) to mark change-of-state events, diagnosis events, coordination, and negation. After we have completed the annotation, we build NLP systems to automatically identify named entities and medical events, which yield an f-score of 94.7% and 91.8%, respectively. Keywords: Annotation, Event Extraction, Named-entity recognition 1. Introduction Clinical notes describe the progress of disease in great level of detail. Although notes individually present a single snapshot in time, the clinical narrative in each note often makes explicit or implicit references to a previous note. To capture this level of information is important for the de- tection of phenotypes/diseases such as pneumonia. For in- stance, in reports that do not explicitly mention that the pa- tient has pneumonia, an important cue for pneumonia de- tection is pneumonia-related symptoms and their change- of-state through time, which is often encoded in sequential ICU progress notes and chest x-ray reports. As an example, the text snippet “No change in diffuse lung disease consis- tent with edema. Patchy bilateral lung consolidation has diminished. No new focal abnormalities” belongs to a chest x-ray note of a patient who is labeled in the gold standard as negative for pneumonia. Our pneumonia prediction system mislabels the patient as positive for pneumonia, because the system primarily used n-grams and UMLS concepts as fea- tures and did not capture change-of-state information in the feature representation (Tepper et al., 2013). To address this problem, in our previous study (Klassen et al., 2014), we proposed to annotate change-of-state events in the text snippets with a tuple schema of five fields ⟨location, attribute, value, change-of-state, reference⟩. In this paper, we discuss the extensions we have made to the schema to add annotation of diagnosis events and to handle coordination and negation of tuple fields. We then report on the experimental results of the NLP systems we build to au- tomatically identify change-of-state and diagnosis events. 2. Creating a Corpus for Phenotype Detection Early detection and treatment of ventilator associated pneumonia (VAP) is important as it is the most com- mon healthcare-associated infection in critically ill patients. Even short-term delays in appropriate treatment for patients with VAP are associated with high mortality rates, longer- term mechanical ventilation, and excessive hospital costs. Our research goal is to build NLP systems which assist healthcare practitioners in identifying patients who are de- veloping critical illnesses (e.g., VAP). Figure 1: A sample chest x-ray report 2.1. PNA/CPIS Detection In our previous study (Tepper et al., 2013), we created an annotated …

Content truncated to 500 words