A mapper-based classifier for patient subgroup prediction

The Mapper algorithm serves as a valuable tool for constructing graph-based representations of high-dimensional data, capturing both topological and geometric information at a specified resolution. This approach has found successful applications in patient subgroup discovery, offering valuable insights from diverse biomedical datasets. Subgroup discovery seeks to identify homogeneous patient subsets within large, heterogeneous cohorts, ultimately enabling more personalized and effective treatment strategies on an individual patient level.

The discovery of new subgroups becomes even more beneficial when we possess an effective method for determining whether a new patient should be classified as a member of a particular subgroup or not. While machine learning methods have proven their utility in various subgroup classification tasks across different medical applications, our experimental evaluations across multiple medical datasets have revealed their challenges in learning patterns associated with small regions of a Mapper graph, resulting in low classification accuracy. To address this issue, we propose a hypothesis that the prediction of subgroups detected via the Mapper graph should rely on Mapper graph-based distances. Consequently, in this study, we introduce a novel approach called the Mapper k Nearest Neighbor algorithm for performing subgroup classification on a Mapper graph. We substantiate the effectiveness of our method through experiments conducted on five real-life gene expression cancer datasets.