Iterative feature selection for classification of platinum-resistant ovarian neoplasms

Abstract

Predicting ovarian cancer patient response to platinum (Pt) based chemotherapy through gene identification is essential for enabling personalized treatment. This study addresses the challenge of limited data and high dimensionality inherent in genomic analyses. We present an iterative feature selection approach using a multinomial Naive Bayes (MNB) classifier trained on mutation annotation format (MAF) data from 31 patients comprised of Pt sensitive and resistant groups. The MNB model is specifically designed for classification problems where the features represent the frequency of occurrence of certain events. The algorithm repeatedly trains MNB models on randomly selected subsets of up to five genes, evaluating performance via cross-validation (CV). By iterating this process across a grid of iteration and CV metrics, we identify genes most frequently selected in top-performing models, revealing a potential genomic signature for platinum sensitivity/resistance. This novel approach leverages machine learning to effectively uncover clinically relevant biomarkers from limited genomic data.