Applying Feature Selection to Improve Predictive Performance and Explainability in Lung Cancer Detection with Soft Computing

Abstract

The field of biomedicine is focused on the detection and subsequent treatment of various complex diseases. Among these, cancer stands out as one of the most studied, due to the high mortality it entails. The appearance of cancer depends directly on the correct functionality and balance of the genome. Therefore, it is mandatory to ensure which of the approximately 25,000 human genes are linked with this undesirable condition. In this work, we focus on a case study of a population affected by lung cancer. Patient information has been obtained using liquid biopsy technology, i.e. capturing cell information from the bloodstream and applying an RNA-seq procedure to get the frequency of representation for each gene. The ultimate goal of this study is to find a good trade-off between predictive capacity and interpretability for the discernment of this type of cancer. To this end, we will apply a large number of techniques for feature selection, using different thresholds for the number of selected discriminant genes. Our experimental results, using Soft Computing techniques, show that model-based feature selection via Random Forest is essential for both improving the predictive capacity of the models, and also their explainability over a small subset of genes.