Predicting with Confidence: A Case-Based Reasoning Framework for Predicting Survival in Breast Cancer

Abstract

There is usually a trade-off between predictive performance and transparency, where the reasoning process behind an algorithm is shielded behind a ”black-box.” In medical domains, experts being responsible for their decisions need to understand the reasons behind machine-generated recommendations. This paper presents a transparent case-based survival analysis framework that automatically retrieves an optimal number of solved survival cases and adapts them to predict the survival of a new case. With this methodology, retrieved and adapted survival cases lend an insight into which cases a prediction is based on. Our framework is capable of integrating DNA methylation, gene expression, and their combination in breast cancer. Additionally, we test our approach with and without feature selection and demonstrate the usefulness of the adaptation phase. We demonstrate that our framework performs at least as effectively as other state-of-the-art methods while affording greater explainability.