B1090
Title: SurvBoard: Standardized benchmarking for cancer survival models
Authors: David Wissel - ETH Zurich (Switzerland) [presenting]
Nikita Janakarajan - IBM Research Europe (Switzerland)
Aayush Grover - ETH Zurich (Switzerland)
Enrico Toniato - IBM Research Europe (Switzerland)
Maria Rodriguez Martinez - IBM Research Europe (Switzerland)
Valentina Boeva - ETH Zurich (Switzerland)
Abstract: Survival analysis represents an important application for clinicians and medical researchers, especially in the cancer setting. Recently, multi-omics data have been widely used in addition to clinical data to stratify patients according to their clinical outcomes with varying levels of success. Despite recent work on benchmarking methods for cancer survival prediction, there is still a need for the standardization of various factors, including the choice of clinical covariates, validation strategies, and factors surrounding cohort selection. A novel benchmark, SurvBoard, is proposed which standardizes these design choices to ensure comparability between cancer survival models of all types. SurvBoard includes 32 cancer datasets from diverse sources, including both multi-omics and clinical-only cohorts. It is shown that while the use of transcriptomic data almost universally helps improve prediction performance, future work is needed to best exploit other omics modalities. The experiments also reveal that covariate interactions do not drive model performance, as additive models rarely underperform models with interactions. In addition, survival models with strong inductive biases, such as the Proportional Hazards model, often perform surprisingly well, even relative to models with significantly fewer assumptions, such as discrete-time methods. Finally, a web service is offered that enables continuous extensions of SurvBoard by other stakeholders.
