We modeled mortality of Intensive Care Unit (ICU) patient using unstructured clinical notes. The mortality of ICU patient is critical for the betterment of patient care since it provides the summarization of a patient’s severity from complex physiological information. Our objective is to make a model that learns a compact representation of clinical text feature space consisted of two different text-derived features; semantically enriched concepts from ontology and topics generated from probabilistic topic modeling. The underlying assumption is that a feature space of richer information can be obtained when the two different feature learning schemes are combined. With these various feature sets, we find a compact representation and classification by sparse group regularization on logistic regression. This representation is more compact and leads better predictability on patient mortality.