Using multiple sentiment dimensions of nursing notes to predict mortality in the intensive care unit

Unstructured clinical data such as nursing notes or discharge summaries are seldom used to predict clinical outcomes, despite containing a lot of information. This study examined several sentiment dimensions of nursing notes for their contributions to 30-day mortality prediction, in the presence of a known predictor of 30-day mortality (SAPS-II). Sentiment dimensions were extracted using a combination of word frequency and machine learning methods. Gender and type of intensive care unit (ICU) were also included as candidate features. The sentiment dimensions are then ranked via a correlation feature selection filter and a recursive feature elimination. SAPS-II was consistently ranked as the best predictor. With a random forest classifier, the predictive performance was significantly improved with sentiment dimensions features (p-value <;0.05) (mean [standard deviation] area under the receiver operating curve with sentiment dimensions: 0.827 [0.011]; without sentiment dimensions: 0.572 [0.010]). Similar improvement was also observed with a logistic regression classifier (p-value <;0.05) (with sentiment dimensions: 0.824 [0.012]; without sentiment dimensions: 0.785 [0.013]). Improvements to mortality prediction is possible by including sentiment analysis.