The invention relates to a strategy for predicting the glycemic profile of a patient with type 1 diabetes (T1D) on the basis of past physiological measurements collected with minimally-invasive and non-invasive on-body sensors and past therapeutic actions. Predictions are long-term and for multiple steps ahead, specifically 30-, 60- and 90-minutes ahead. The chore of the method lies in the prediction model which is comprised of a Convolutional Neural Network (CNN) followed by a Long Short Term Memory (LSTM) network. In our proposed architecture, multiple layers of convolutional blocks are used for feature extraction, while the LSTM blocks are used for learning the temporal dynamics.