Investigating the prognostic accuracy of standardized data mining algorithms in intensive care unit

Objectives: Modern clinicians use scalable data mining models to evaluate their hypotheses. The purpose of this paper is to present the lessons learned in solving prognostic problems in Intensive Care Unit (ICU) by using data mining models developed with standardized algorithms as an alternative solution to clinical assessment tools. Methods: The study included data from 201 ICU patients (156 male and 45 female) that were assessed by means of the APACHE II, the SOFA and the ISS as well as free thyroxin fT4, total triiodothyronine (TT3) T3, thyrotropin (TSH), corticotropin (ACTH), prolactin, cortisol and dehydroepiandrosterone sulphate (DHEAS) and the Synacthen test. We formulated three data mining models - a decision tree (DTM), a neural network (NNM), and a linear regression (LRM)- using the standardized algorithms of Microsoft™ SQL Server 2005 Data Mining Platform. The outcomes were compared against those of ICU clinical assessment tools and hormone measurements. Results: From the ROC plot analysis the APACHE II score was only marginally better than the SOFA or ISS score in predicting ICU survival. Moreover, the standardized data mining models applied on endocrine parameters were not outperformed by the APACHE II, SOFA or ISS scores alone in predicting ICU survival. Conclusions: From negative results, useful information can always be deduced. Our results point to the need to use custom algorithms to support particular ICU mining needs in lieu of standardized algorithms.