A Rank-Consensus Ensemble Feature Selection and Soft-Voting Framework for Interpretable Breast Cancer Prediction

Abstract

Early and correct prediction of breast cancer is clinically relevant due to the complex, correlated and non-linear nature of patterns in fine-needle aspiration measurements. In this work, a rank consensus ensemble feature selection and soft-voting classification approach to classify the Breast Cancer Wisconsin Diagnostic data sets into distinction between Malignant and Benign classes is presented. The idea behind the proposed framework is to combine three complementary flavours of feature-selection: dependency in non-linear manner using MI, separability of classes using Fisher score and recursive feature elimination aimed at example-model guided relevance. The outputs of these selectors are then normalized and the score combined to create a consensus score and feature importance score for reduced redundancy and a small feature-set that can be clinically interpreted. Logistic regression, random forest and decision tree learners are then combined in a probabilistic soft-voting manner to enhance the stability of prediction. The most informative features found were: worst radius, worst concave points, mean concave points, mean radius and worst perimeter. Results from the hold-out evaluation set showed that the proposed model was able to correctly classify 109 out of 114 cases with 95.61% accuracy, 95.24% precision, 93.02% sensitivity, 97.18% specificity and 94.12% F1 score. Unconfined analysis of the confusion matrix also reveals low false-positive and false-negative rates showing this proposed framework offers good balance of clinical sensitivity and diagnostic specificity. The key novelty this will bring is that the pipeline is interpretable, it is low dimensional and ensemble based, and it will help computer aided screening of breast cancer, which also is feasible in the computational space of clinical decision support.