This research aims to improve the efficiency of intensive care units (ICU) by improving patient flow. A UK ICU provides a case study for this research. Of particular interest in this work is the impact of 'late admissions', which account for 13.8% of all first-time admissions to this ICU. Patients admitted to the ICU more than a day after entering the hospital are shown to have higher mortality rates and to stay longer in the ICU. Mortality and length of stay (LoS) are predicted to assist ICU modelling. After comparing different binary prediction models, three sets of logistic regression models have been built to predict patients' mortality from different admission groups (such as planned, unplanned, late or re-admission). The overall performance of the prediction models developed in this project is better than using ICNARC (Intensive Care National Audit and Research Centre) probability directly. LoS of individuals is found to be hard to predict. A new method for modelling LoS is tested and applied. LoS is split into three sub-parts, admission hour, nights spent in the ICU and discharge hour, for which empirical distribution functions are used. We describe a Discrete Event Simulation (DES) model to investigate the impact of the late admission group and strategies for improving efficiency by bringing patients into the ICU earlier. Mortality prediction models and the new method of LoS modelling are incorporated into the DES input distributions. Several scenarios are investigated including varying resource number and earlier admission of patients. A key finding is that the ICU can accommodate 20% more unplanned patients based on the current situation if the late admission group can be reduced to 5% of all first-time admissions. We also consider an epidemic scenario: it is demonstrated that the ICU would only be able to cope with a mild epidemic.