In an intensive care unit (ICU), management of blood glucose (BG) levels is essential to avoid adverse outcomes, including death. Traditional machine learning (ML) and deep learning (DL) methods have previously been used for BG level forecasting. However, these have mostly focused on data from populations with diabetes outside of an ICU setting. In addition, these are often inaccurate when making forecasts in the hypo- and hyperglycaemic ranges due to the relative infrequency of such data instances. Using data derived from the MIMIC-III clinical database, this thesis investigates the impact of two data augmentation (DA) methods on the accuracy of BG level forecasting in an ICU setting. The first DA method relies on random data transformations, with the second relying on using a state-of-the-art generative adversarial network (GAN) to create synthetic data. The effect of these DA methods on a predictive model’s forecasting accuracy is investigated for traditional ML and DL models (XGBoost and LSTM, respectively), paying particular attention to the hyper- and hypoglycaemic BG ranges. The findings of this thesis could potentially contribute towards improved clinical outcomes of patients in an ICU by providing accurate early warnings of dysglycaemia, and reducing the workload of ICU staff.