Timely identification of hospital inpatients at risk of deterioration facilitates interventions to support their recovery. Many hospitals implement early warning scores to detect abnormal patient vital signs, such as the National Early Warning Score 2 (NEWS2). However, these are typically based on a snapshot of the most recent vital signs, rather than exploiting trends over time that clinical intuition suggests may also be informative. Multiple approaches, including recently described methods, have been developed to predict patient deterioration from time series. We therefore compared the effectiveness of different mortality prediction models, including clinical scoring systems, classical machine learning models and state-of-the-art deep learning models using both snapshot and time series vital sign data. No significant improvement in model performance was observed using predictions from time series compared to using the last observation of the time series and non-temporal features such as demographics. Our study comprehensively compares different model types, and provides recommendations for developing predictive models and guidance for what evaluation is needed before considering deploying such models in inpatient care.