Improving structural medical process comparison by exploiting domain knowledge and mined information

Process model comparison and similar process retrieval is a key issue to be addressed in many real-world situations, and a particularly relevant one in medical applications, where similarity quantification can be exploited to accomplish goals such as conformance checking, local process adaptation analysis, and hospital ranking. In this paper, we present a framework that allows the user to: (i) mine the actual process model from a database of process execution traces available at a given hospital; and (ii) compare (mined) process models. The tool is currently being applied in stroke management.

Our framework relies on process mining to extract process-related information (i.e., process models) from data. As for process comparison, we have modified a state-of-the-art structural similarity metric by exploiting: (i) domain knowledge; (ii) process mining outputs and statistical temporal information. These changes were meant to make the metric more suited to the medical domain.

Experimental results showed that our metric outperforms the original one, and generated output closer than that provided by a stroke management expert. In particular, our metric correctly rated 11 out of 15 mined hospital models with respect to a given query. On the other hand, the original metric correctly rated only 7 out of 15 models. The experiments also showed that the framework can support stroke management experts in answering key research questions: in particular, average patient improvement decreased as the distance (according to our metric) from the top level hospital process model increased.

The paper shows that process mining and process comparison, through a similarity metric tailored to medical applications, can be applied successfully to clinical data to gain a better understanding of different medical processes adopted by different hospitals, and of their impact on clinical outcomes. In the future, we plan to make our metric even more general and efficient, by explicitly considering various methodological and technological extensions. We will also test the framework in different domains.