This paper proposes a formal optimisation framework and algorithms for data-aware process mining with event duplication that relaxes the usual one-event-label-one-process-model-node restriction. We put forward a hierarchical representation of the event attribute values and event labelling to achieve the best balance of the complexity and precision of the process model. We posit a new quality measure, relevance, which measures how well and how precisely a process model matches a given event log. The process model optimisation consists of determining (i) the process model with labels and attribute values for each node and transition functions for each arc and (ii) the event game stipulating how each trace of the event log is played in the process model. This article also proposes a dynamic programming algorithm for optimising event games, an exact method for optimal setting of node attributes and arc transition functions, and heuristic algorithms for process model optimisation. Numerical results show the efficiency of the algorithms with respect to relevant benchmarks and an 18% improvement in the model relevance. Applications on sarcoma care pathways reveal their dependency on attributes such as surgery quality and tumour size. Our approach clearly shows how both care event repetition and data impact sarcoma care pathways whereas other data-aware miners fail.