Expectation–Maximization (EM) Clustering as a Preprocessing Method for Clinical Pathway Mining

Hospital information systems (HIS) are service-oriented systems that focus on payment for medical services. Because all HIS coding for diseases and clinical processes are payment-oriented, they may differ from clinicians’ concepts of diseases and processes. HIS in large-scale hospitals in Japan utilize Diagnostic Procedure Combination (DPC) codes, a disease-coding system that focuses on the use of medical resources. Although DPC codes are very precise for diseases requiring surgery, such as cataracts and lung cancer, classification codes for diseases that do not require surgery, such as cerebral infarction, are less precise, with a single category often covering many subtypes with different clinical courses. This paper proposes a preprocessing method that splits DPC codes into subgroups prior to the application of dual clustering-based clinical pathway mining. This method applies expectation–maximization (EM) clustering to the length of patient stay in the hospital using Akaike Information Criteria (AIC) to select the number of clusters. A dual mining method is subsequently applied to the datasets of subgroups and the meanings of subtype clusters are explored using a text mining method. The proposed method was evaluated using datasets from an HIS at Shimane University hospital as preprocessing for clinical pathway mining. The experimental results showed that the proposed method correctly generated subgroups from the more generalized DPC codes and that the clinical pathways identified after this preprocessing capture the characteristics of processes in real clinical settings.