Abstract
BACKGROUND: Advanced data mining techniques such as decision trees have been successfully used to predict a variety of outcomes in complex medical environments. Furthermore, previous research has shown that combining the results of a set of individually trained trees into an ensemble-based classifier can improve overall classification accuracy. This paper investigates the effect of data pre-processing, the use of ensembles constructed by bagging, and a simple majority vote to combine classification predictions from routine pathology laboratory data, particularly to overcome a large imbalance of negative Hepatitis B virus (HBV) and Hepatitis C virus (HCV) cases versus HBV or HCV immunoassay positive cases. These methods were illustrated using a never before analysed data set from ACT Pathology (Canberra, Australia) relating to HBV and HCV patients. RESULTS: It was easier to predict immunoassay positive cases than negative cases of HBV or HCV. While applying an ensemble-based approach rather than a single classifier had a small positive effect on the accuracy rate, this also varied depending on the virus under analysis. Finally, scaling data before prediction also has a small positive effect on the accuracy rate for this dataset. A graphical analysis of the distribution of accuracy rates across ensembles supports these findings. CONCLUSIONS: Laboratories looking to include machine learning as part of their decision support processes need to be aware that the infection outcome, the machine learning method used and the virus type interact to affect the enhanced laboratory diagnosis of hepatitis virus infection, as determined by primary immunoassay data in concert with multiple routine pathology laboratory variables. This awareness will lead to the informed use of existing machine learning methods, thus improving the quality of laboratory diagnosis via informatics analyses.