Data mining from a patient safety database: the lessons learned

The issue of patient safety is an extremely important one; each year in the UK, hundreds of thousands of people suffer due to some sort of incident that occurs whilst they are in National Health Service care. The National Patient Safety Agency (NPSA) works to try to reduce the scale of the problem. One of its major projects is to collect a very large dataset, the Reporting and Learning System (RLS), which describes several million of these incidents. The RLS is used as the basis for research by the NPSA. However, the NPSA has identified a gap in their work between high-level quantitative analysis and detailed, manual analysis of small samples. This paper describes the lessons learned from a knowledge discovery process that attempted to fill this gap. The RLS contains a free text description of each incident. A high dimensional model of the text is calculated, using the vector space model with term weighting applied. Dimensionality reduction techniques are used to produce the final models of the text. These models are examined using an anomaly detection tool to find groups of incidents that should be coherent in meaning, and that might be of interest to the NPSA. A three stage process is developed for assessing the results. The first stage uses a quantitative measure based on the use of planted groups of known interest, the second stage involves manual filtering by a non-expert, and the third stage is assessment by clinical experts.