Physiological models represent a useful form of knowledge, but are both difficult and time consuming to generate by hand. Further, most physiological systems are incompletely understood. This paper addresses these two issues with a system that learns qualitative models from physiological signals. The qualitative representation of models allows incomplete knowledge to be encapsulated, and is based on Kuipers’ approach used in his QSIM algorithm. The learning algorithm allows automatic generation of such models, and is based on Coiera’s GENMODEL algorithm. We describe experiments using the learning system on data segments obtained from six patients during cardiac bypass surgery. Useful model constraints were obtained, representing both general physiological knowledge and knowledge particular to the patient being monitored. The effects of varying parameters for front-end signal processing, and varying fault tolerance levels for the subsequent learning by GENMODEL are also discussed.