Computing the Hazard Ratios Associated With Explanatory Variables Using Machine Learning Models of Survival Data

PURPOSE The application of Cox proportional hazards (CoxPH) models to survival data and the derivation hazard ratio (HR) are well established. Although nonlinear, tree-based machine learning (ML) have been developed applied analysis, no methodology exists for computing HRs associated with explanatory variables from such models. We describe a novel way compute ML using SHapley Additive exPlanation values, which is locally accurate consistent quantify variables’ contribution predictions. METHODS used three sets publicly available consisting patients colon, breast, or pan cancer compared performance CoxPH state-of-the-art model, XGBoost. To HR XGBoost values were exponentiated means over two subgroups was calculated. CI computed via bootstrapping training generating model 1,000 times. Across sets, we systematically all variables. Open-source libraries in Python R analyses. RESULTS For colon breast comparable, showed good consistency HRs. In pan-cancer set, agreement most but also an opposite finding between result. Subsequent Kaplan-Meier plots supported model. CONCLUSION Enabling can help improve identification risk factors complex enhance prediction clinical trial outcomes.