Avoidance and mitigation of disruptions are crucial problems in ITER and are becoming increasingly relevant at JET with the installation of the new ITER-Like Wall (ILW). In order to do so, disruption prediction capability is a pre-requisite. This work summarizes the evolution of our research lines regarding disruption prediction at JET during the last years and discusses their potential applicability to ITER. Disruption predictors are typically based on binary classification techniques. Given a dataset of N examples represented by pairs (xi, yi, i = 1,...N, (where xi Rm is a vector of dimension m that represents features of distinctive nature among the N examples and yi {+1,­1} is the label about the plasma behaviour, i.e. disruptive or non-disruptive), a training process determines a mathematical function to split the feature space into two regions. After completing the training, new feature vectors xnew Rm are classified as disruptive or non-disruptive depending on the parameter space region where they are located. We have tackled three main research topics about disruption prediction in JET from 2008: advanced predictors, advanced predictors from scratch and disruption time predictors. All of them are based on the fulfilment of four important requirements. Firstly, they use multidimensional data. Secondly, they are based on discovering relations between signals to identify a forthcoming disruption. Thirdly, they have been programmed to provide deterministic responses. Last but not least, they were designed to use only real-time signals.