For many years, machine learning tools have proved to be very powerful disruption predictors in tokamaks. On the other hand, the vast majority of the techniques deployed assume that the input data is independent and is sampled from exactly the same probability distribution for the training set, the test set and the final real time deployment. This hypothesis is certainly not verified in practice, since the experimental programmes evolve quite rapidly, resulting typically in ageing of the predictors and consequent suboptimal performance. This paper describes various adaptive training strategies that have been tested to maintain the performance of disruption predictors in non-stationary conditions. The proposed approaches have been implemented using new ensembles of classifiers, explicitly developed for the present application. The improvements in performance are unquestionable and, given the difficulties encountered so far in translating predictors from one device to another, the proposed adaptive methods from scratch can therefore be considered a useful option in the arsenal of alternatives envisaged for the next generation of devices, particularly at the very beginning of their operation.