Nurse notes often contain subjective information of a patient's health status. However, these have not been widely used to predict clinical outcomes. Advances in natural language processing enable to extract information from unstructured clinical documents such as nurse notes. In this paper, we use TF-IDF representations of nurse notes to predict Intensive Care Unit (ICU) mortality while controlling for other candidate features such as gender, ICU type, age of patient at first admission, SAPS (Simplified Acute Physiology Score) II score, SAPS II probability, polarity and subjectivity scores of each nurse note. We introduce an ensemble of CNN and MLP model to predict 30-day ICU mortality. We apply our model to MIMIC III which is a medical benchmark dataset. We use TF-IDF representation of nurse notes as input to CNN and other ICU features to MLP network. Experimental results demonstrate that proposed ensembled model with nurse notes have higher performance than standalone models.