Every year, large numbers of patients in National Health Service (NHS) care suffer because
of a patient safety incident. The National Patient Safety Agency (NPSA) collects large
amounts of data describing individual incidents. As well as being described by categorical
and numerical variables, each incident is described using free text.
The aim of the work was to find quite small groups of similar incidents, which were of
types that were previously unknown to the NPSA. A model of the text was produced, such
that the position of each incident reflected its meaning to the greatest extent possible.
The basic model was the vector space model. Dimensionality reduction was carried
out in two stages: unsupervised dimensionality reduction was carried out using principal
component analysis, and supervised dimensionality reduction using linear discriminant
analysis. It was then possible to look for groups of incidents that were more tightly packed
than would be expected given the overall distribution of the incidents.
The process for assessing these groups had three stages. Firstly, a quantitative measure
was used, allowing a large number of parameter combinations to be examined. The groups
found for an ‘optimum’ parameter combination were then divided into categories using a
qualitative filtering method. Finally, clinical experts assessed the groups qualitatively.
The transition probabilities model was also examined: this model was based on the
empirical probabilities that two word sequences were seen in the text.
An alternative method for dimensionality reduction was to use information about the subjective meaning of a small sample of incidents elicited from experts, producing a mapping
between high and low dimensional models of the text.
The analysis also included the direct use of the categorical variables to model the incidents,
and empirical analysis of the behaviour of high dimensional spaces.