TCPM: Topic-Based Clinical Pathway Mining

Clinical pathway is important for improving medical quality, reducing cost and regulating resource. However, a static, non-adaptive clinical pathway designed by experts with limited data can be hardly implemented in practice. Thus, mining the execution clinical pathway from various historical data is meaningful. Existing works focus on applying either process mining or clustering methods on medical data. These methods generally produce low-granularity process models or unordered trace groups with similar treatment behaviors. In this paper, we propose a topic-based clinical pathway mining approach, which is concise, interpretable and of sequential information. We start from billing data, and use Latent Dirichlet Allocation to cluster billing items without specifying the topic number. The treatment of each day is represented as a set of topics, which convey the treatment goals. To emphasize critical and essential activities, we prune the low-frequency topics and remove sub-traces. Finally, by applying fuzzy mining method on these topic sequences, we can discover the execution clinical pathway. The experiments on a real-world data set show the effectiveness and practicability of our approach.