Comparison of three classifiers for breast cancer outcome prediction

Predicting the outcome of cancer is a challenging task; researchers have an interest in trying to predict the relapse-free survival of breast cancer patients based on gene expression data. Data mining methods offer more advanced approaches for dealing with survival data. The main objective in cancer treatment is to improve overall survival or, at the very least, the time to relapse ("relapse-free survival"). In this work, we compare the performance of three popular interpretable classifiers (decision tree, probabilistic neural networks and Naïve Bayes) for the task of classifying breast cancer patients into recurrence risk groups (low or high risk of recurrence within 5 or 10 years). For the 5-year recurrence risk prediction, the highest prediction accuracy was reached by the probabilistic neural networks classifier (Acc = 76.88% ± 1.09%, AUC=77.41%). For the 10-year recurrence risk prediction, the decision tree classifier and the probabilistic neural networks presented similar prediction accuracies (70.40% ± 1.36% and 70.50% ± 1.13%, respectively). However, while the PNN classifier achieved this accuracy using only 10 features with the highest information gain, the decision tree classifier needed 100 features to achieve comparable accuracy and its AUC was significantly lower (66.4% vs. 77.1%).