Developing significant predictive models in critical illness

1 October, 2004 to 30 November, 2007


  • Hendrik Blockeel
  • Maurice Bruynooghe
  • Jan Ramon

Current risk prediction models in critical illness are empirically derived and based on rather traditional statistical methods, e.g. logistic regression or multivariate analysis. Their performance is good in comparing severity of disease between different patient populations, but poor in predicting behaviour of individual patients. In patients on an intensive care ward, large amounts of time-variant and non-time-variant data are almost continuously registered and digitally archived. This large database can be structured and analysed using relational data mining techniques. Hidden relationships between parameters can be revealed. Complex biological systems and their individual behaviour can be predicted using compact dynamic data based modelling. Combining these two techniques, predictive models both at the level of the individual patient and the patient population will be developed. Further research will be done regarding the clinical relevance of the parameters derived from this model.