Sequential pattern mining attempts to extract frequent behaviors from a sequential dataset. When sequences are labeled, it is interesting to extract behaviors that characterize each sequence class. This task is called discriminant pattern mining. In this paper, we introduce discriminant chronicle mining. Conceptually, a chronicle is a temporal graph whose vertices are events and whose edges represent numerical temporal constraints between these events. We propose DCM, an algorithm that mines discriminant chronicles. It is based on rule learning methods that extract the temporal constraints. Computational performances and discriminant power of extracted chronicles are evaluated on synthetic and real data. Finally, we apply this algorithm to the case study consisting in analyzing care pathways of epileptic patients.