Identifying Documentation of Delirium in Clinical Notes through Topic Modeling

Delirium is prevalent, costly and under detected. Diagnosis or classification is largely done in text and often the narrative is behavioral, fuzzy and informal. To enable just-in-time decision support, we set out to identify the documentation of delirium in clinical notes. Two experts annotated documents from a Pittsburgh dataset. We experimented with 3 different topic modeling methods including LDA and 2 ICD-based methods and a keyword search method for the identification of delirium related documents and sentences in clinical notes. As expected, the keyword search method is highly specific but insufficiently sensitive when searching for mentions of delirium in the documents. All 3 topic models performed better in terms of recall but worse in precision when compared with keyword search. The ICD-2 method, in particular, achieved a F-score of 0.677. In contrast, the keyword search reached a F-score of 0.442. Implications regarding decision support design, enhancing collaboration within clinical teams and improving resource utilization were discussed.