Various methods for predicting outputs in an icon-based pediatric communication aid are explored. These include a time-based frequency counter, statistical analysis based on previous selections, and feed-forward neural networks. In order to properly train the prediction mechanisms, a large amount of data are required. This thesis develops and evaluates two methods for data generation when little or no labeled communication data is available. The first method is to use a family member's knowledge of the communication patterns of the patient, and the second is to use the dependencies found during the patient's initial device usage to generate additional sequences. With a single subject in critical care, it was found that a neural network predictor based on a backpropagation algorithm achieved the highest prediction rate. The minimum amount of empirically collected data required to generate useful training sequences, and the minimum length of generated sequences to adequately train the predictors are also investigated.