Abstract: Since 2012, the High Authority for Health has imposed a new paradigm: "Never a first time on the patient" which has led to the development of medical simulation. We present SVP-OR, a generator of reactive scenarios designed to provide digitally assisted training for interns in anaesthesiology (SVPOR users). To simulate a virtual patient (VP), we predict the short-term evolution of the VP's multivariate time series grown so far, each time an action is triggered by the intern or the virtual medical team. The prediction problem is tackled as a case-based reasoning approach: the evolution of the VP is computed from the real patients showing some region of their histories similar to the VP's recent history. A grammar built from real patients' action traces drives the consistent scheduling of the SVP-OR user-induced actions and of the medical team's virtual actions. Our contributions are the design of a contextualized multidimensional pattern recognition approach, and a vast comparative study of dissimilarity measures focused on short time series. We evaluated four variants of our generic SVP-OR framework. We showed in all cases that SVP-OR is able to generate on-the-fly realistic predictions.

Keywords: Computer-assisted Medical Training; Virtual Patient; Operating Room ; Multivariate Time Series Prediction; Event Trace;