Risk factors and prediction of very short term versus short/intermediate term post-stroke mortality: a data mining approach

Abstract

Data mining and knowledge discovery as an approach to examining medical data can limit some of the inherent bias in the hypothesis assumptions that can be found in traditional clinical data analysis. In this paper we illustrate the benefits of a data mining inspired approach to statistically analysing a bespoke data set, the academic multicentre randomised control trial, U.K Glucose Insulin in Stroke Trial (GIST-UK), with a view to discovering new insights distinct from the original hypotheses of the trial. We consider post-stroke mortality prediction as a function of days since stroke onset, showing that the time scales that best characterise changes in mortality risk are most naturally defined by examination of the mortality curve. We show that certain risk factors differentiate between very short term and intermediate term mortality. In particular, we show that age is highly relevant for intermediate term risk but not for very short or short term mortality. We suggest that this is due to the concept of frailty. Other risk factors are highlighted across a range of variable types including socio-demographics, past medical histories and admission medication. Using the most statistically significant risk factors we build predictive classification models for very short term and short/intermediate term mortality.

Keywords: Data mining; Medical relevance; Mortality; Naïve Bayes analysis; Prediction; Risk factors; Stroke.