In my PhD thesis I developed time-series predictive models for lung tumor radiation therapy. A multilayer perceptron (MLP) with one hidden layer and a quadratic neural unit (QNU) were developed as predictive models. The adaptation rule for the models was batch optimization (Levenberg–Marquardt) implemented with a retraining technique on a sliding window, that is, a moving window that retrains the model for every new sample.
The thesis demonstrated the predictive capability of the models with the moving window on periodic and highly nonlinear respiratory time series, predicting real lung-movement data. The prediction results met the goal of the work: an accuracy of 1 mm with a prediction horizon of up to 1 second.