Understanding Inequalities in Cancer Survival Using Bayesian Machine Learning

• 成果类型: Article; Early Access
• 署名作者: Basak, Piyali; Maringe, Camille; Rubio, F. Javier; Linero, Antonio R.
• 署名单位: Merck & Company; Merck & Company USA; University of London; London School of Hygiene & Tropical Medicine; University of London; University College London; University of Texas System; University of Texas Austin
• 刊物名称: JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
• ISSN/ISSBN: 0162-1459
• DOI: 10.1080/01621459.2025.2547968
• 发表日期: 2025
• 关键词: regression-models; diagnosis; inference

摘要

Most cancer patients are diagnosed after the age of 60, often with existing chronic health conditions (comorbidities), that can delay diagnosis and complicate treatment, prognosis, and monitoring. These comorbidities may exacerbate existing sociodemographic inequalities in cancer survival. While much research has focused on how comorbidities affect overall survival, national and international institutions typically prefer the relative survival framework for population-based studies. This framework decomposes an individual's overall hazard into a known population hazard and an excess hazard attributable to cancer. Estimands derived from the excess hazard, such as net survival, are widely used to assess interventions and inform policy. In this article, we use a Bayesian machine learning approach to estimate the excess hazard and identify vulnerable subgroups with a higher excess hazard, using Bayesian additive regression trees (BART). We develop a proportional hazards version of BART for the relative survival context and extend it to accommodate nonproportional hazards. We also provide tools for model interpretation and posterior summarization. This is applied to colon cancer data from England to provide insights when paired with state-of-the-art data linkage methods. We then identify drivers of inequalities in cancer survival through variable importance quantification. Supplementary materials for this article are available online, including a standardized description of the materials available for reproducing the work.