Analysis of decision tree performance in predicting the relationship between a scored outcome and multiple single nucleotide polymorphisms

We describe the results of an experiment for which we used decisional tree analysis to find a relationship between multiple single nucleotide polymorphisms (SNPs) and a scored health outcome based on a given dataset. We used Gini Index and Minimum Description Length (MDL) to optimize the accuracy of predicting the scored health outcome using SNPs as input variables. Decision tree models were generated using different configuration of pruning, sampling and split functions. We analyzed two parameters, the accuracy of the prediction and the Cohen's kappa coefficient, for each generated model. In regard to scoring stage, the experiment was divided in two phases. The first phase used a more detailed dataset and the second phase a less detailed dataset. The best models, with the highest accuracy, were obtained for a decision tree that used the split function Gini Index with the MDL pruning, and with the random sampling method. The model with the best prediction accuracy had a 91% identification rate. The results obtained for the Cohen's kappa coefficient were less accurate.