Towards a Mixed Learning Strategy for Discovering New Gene Signatures in Breast Cancer Prognosis

This work focuses on developing a mixed-learning method that combines a filter-based metaheuristic searcher with a shallow learning classifier to reduce the feature space while maximizing the breast cancer prognosis classification. The searcher used a genetic algorithm together with the average symmetrical uncertainty (aSU) and ReliefF (aReliefF) filter functions. This modification allowed us to measure the relevance per capita of a group of features (genes). The proposed method was validated on a data set with 396 instances. The most effective classification scheme emerged from the random forest model, utilizing 60 tree predictors and employing the aReliefF objective function. This configuration achieved an average area under the receiver operating characteristic curve (AUC) score of 0.854 and 0.874 for the training and test stages, respectively. Thus, this classification scheme is the best breast cancer prognosis classification strategy. In addition, we identified a set of master genes through the intersection of both objective functions regarding feature relevance. Nevertheless, evaluating this subset in the test set using the top-performing classification scheme yielded a comparatively lower performance (AUC=0.829), underscoring the necessity for additional genes to maximize classification effectiveness.