Unstructured (narrative) data such as nursing notes aren't usually included in predictions of clinical outcomes in ICU patients, although they have been found to provide valuable information. To determine whether they could be used in assessing ICU patients' risk of death, researchers used sentiment analysis, a method of quantifying subjective properties of written text, to evaluate the association between sentiments (impressions or attitudes) in nursing notes and 30-day mortality in ICU patients. A sentiment analysis algorithm was applied to nursing notes extracted from an intensive care database, and sentiment was correlated with 30-day mortality; the researchers controlled for sex, type of ICU, and Simplified Acute Physiology Score II (SAPS II) scores. Each nursing note was assigned a "sentiment polarity" score, as measured using a scale of positivity and negativity, and a "sentiment subjectivity" score, which showed the amount of subjective sentiment used by the nurse.

Among the more than 27,000 ICU patients included in the study, the overall 30-day mortality was 11%. Nursing notes in the electronic health records of patients who survived had significantly higher mean sentiment polarity scores, meaning they were more "positive," than notes in the records of patients who died. Even in the presence of known predictors of 30-day mortality, including the SAPS II score, mean sentiment polarity scores were significant predictors and led to improved prediction accuracy. Survival was also positively correlated with sentiment polarity quartiles, the first quartile having the worst survival and the fourth quartile having the highest survival.

According to the authors, the findings suggest that incorporating unstructured clinical notes into clinical outcome prediction models could lead to better prediction of clinical outcomes.-KR

REFERENCE