Multidimensionales Process-Mining für die Analyse medizinischer Versorgungsprozesse

Process mining allows for the automatic discovery of process models from so-called event logs which capture events that occurred during process execution. Besides traditional business processes, it can also be used to analyze medical treatment processes. Especially the individual characteristics of the patients turn out to be very challenging, because they significantly influence the process behavior. The treatment may be influenced by many features like age, sex, comorbidity and other – possibly nontransparent – features of a patient. However, established process mining techniques consider the entire event log and create a single process model reflecting the treatment of all patients which may blur or even hide the influence of the individual characteristics. To analyze the influence of these characteristics, it is desirable to group patients with similar characteristics and to discover a separated process model for each group of patients. Then, these models can be compared. As the grouping of the patients may depend on the underlying question of the analysis, an approach is required to define the grouping criteria in a simple and flexible way. This work presents a novel approach that considers the patients' characteristics as dimensions forming a multidimensional data space. This so-called multidimensional process mining allows to map the event logs to a data cube which can be queried using OLAP operators. This way, a set of cells can be defined, each representing a group of patients. For each group, a separated process model is discovered. An optional step of consolidation allows for an automatic pre-selection of relevant process models. Finally, the models can be visualized and compared. This approach is implemented as a prototype to show its feasibility. The applicability of the approach to the healthcare domain is demonstrated during a case study in health services research. Additional experiments show that this concept provides a significantly better performance than the state-ofthe-art approach for multidimensional process mining and is scalable with the number of events.