Clinical Notes Mining For Post Discharge Mortality Prediction

Unstructured clinical data such as nursing notes are relatively less explored for building a predictive model for post-discharge mortality despite containing rich information about the health of the patients. Our work examines a simple bag of words (BOW) approach for 7/30/180/365 day mortality prediction of a patient. We have also explored syntactic sentiment dimensions from nursing notes as a predictor of mortality and report the survival analysis results. Our simple BOW model using logistic regression achieved 0.884 AUC for 30-day mortality. We also found out that the polarity may serve as a proxy for patient's post-discharge survival.