Most existing studies used logistic regression to establish scoring systems to predict intensive care unit (ICU) mortality. Machine learning-based approaches can achieve higher prediction accuracy but, unlike the scoring systems, frequently cannot provide explicit interpretability. We evaluated an interpretable ICU mortality prediction model based on Recurrent Neural Networks (RNN) with long short-term memory(LSTM)units. This model combines both sequential features with multiple values over the patient's hospitalization (e.g. vital signs or laboratory tests) and non-sequential features (e.g. diagnoses), while identifying features that most strongly contribute to the outcome. Using a set of 4,896 MICU admissions from a large medical center, the model achieved a c-statistic for prediction of ICU mortality of 0.7614 compared to 0.7412 for a logistic regression model that used the same data, and identified clinically valid predictors (e.g. DNR designation or diagnosis of disseminated intravascular coagulation). Further research is needed to improve interpretability of sequential features analysis and generalizability.