Telemedicine is a new approach of medical practice that is expected to be one of the answers for facing the challenge of chronic diseases management. Development of remote medical surveillance at home relies on our capacity to interpret a growing amount of collected data. In this thesis, we are interested in defining the function that connects the state of the patient to the data given by the different sensors. The main difficulty comes from the uncertainty when assessing the state of the patient. The only reference available is the one that can be given by the medical doctor. We show in this thesis that stochastic modelling and more specifically graphical bayesian formalism allows to treat this question in three ways. The first one consists in representing explicitly the medical expertise. This approach is adapted to the cases in which data is not accessible, and as a consequence, where it is necessary to model directly the diagnosis rules. The second approach that we study is the automatic learning of model parameters that can be performed when enough information is available concerning the expected outputs of the system. Finally, we propose the use of reinforcement for learning medical actions from the observation of the human expert in its everyday practice. Considering the specificity of the medical domain, we study the likelihood criterion for learning an efficient representation of the state space.