SurvivalLVQ: Interpretable supervised clustering and prediction in survival analysis via Learning Vector Quantization

Abstract

Identifying subgroups with similar survival outcomes is a pivotal challenge in survival analysis. Traditional clustering methods often neglect the outcome variable, potentially leading to inaccurate representation of risk profiles. To address this, we present SurvivalLVQ, a novel interpretable method that adapts Learning Vector Quantization (LVQ) to survival analysis. Unlike traditional classification uses of LVQ, SurvivalLVQ groups individuals by survival probabilities and assigns a unique survival curve to each cluster, representing the collective survival behavior within that group. Moreover, it can predict individual survival curves using weighted averages from nearby clusters. When tested on 76 benchmark datasets, it outperformed other clustering methods and showed competitive prediction performance. SurvivalLVQ bridges the gap between clustering techniques and outcome-oriented methods. Its strong clustering performance, coupled with competitive prediction capabilities and with easy to interpret outcomes, make it a promising tool for various applications within survival analysis.