Clustering-based methodology for analyzing near-miss reports and identifying risks in healthcare delivery

Near-miss reports are qualitative descriptions of events that could have harmed patients but did not due to a timely intervention or a convenient evolution of the circumstances. Near-miss reporting has increasingly become a relevant tool to support patient safety efforts since they provide some evidence of risk in the system before patients suffer adverse consequences. Near-misses are usually classified into pre-specified categories that correspond to sources of risk in the system or its processes. Their analysis often consists of tallying classified near-misses to determine risk priorities based on frequency within each pre-specified risk category. Our research aims to use different combinations of near-miss reports to find potential sources of risk. We propose an unsupervised bisecting k-prototypes algorithm for clustering coded near-miss reports to identify relationships between events that would not otherwise have been easily identified. Subsequent study of resulting clusters will lead to the identification of potentially dangerous, but unsuspected system interactions. We illustrate or methodology with preliminary results of its implementation at the University of South Florida Health clinics.