Discovering Interpretable Medical Workflow Models

Medical workflow discovery and analysis can help teams better understand their practice and potentially improve patient outcomes. In the past, medical experts designed hand-made workflow models for this purpose. These models usually needed iterative revision for the experts to reach consensus. Actual practice, however, often deviates from a perceived ideal workflow model. Automatic workflow discovery algorithms have recently been proposed to learn a workflow model from observed activity or event traces. These algorithms, however, can produce spaghetti-like graphical models, with many branches and loops. Although model interpretability is an initial requirement of medical process analysis, it is often difficult or impossible for medical experts to extract knowledge from these chaotic models. We present an algorithm for discovering interpretable medical workflow models that addresses these limitations. We show our preliminary results and evaluate our approach on a real-world medical process.