Evaluation of machine-learning models to measure individualized treatment effects from randomized clinical trial data with time-to-event outcomes

In randomized clinical trials, regression models can be used to explore the relationships between patients' variables (e.g., clinical, pathological or lifestyle variables, and also biomarker or genomics data) and the magnitude of treatment effect. Our aim is to evaluate the value of flexible machine learning models that can incorporate interactions and nonlinear effects of high-dimensional data to estimate individualized treatment recommendations in the setting of such trials with time-to-event outcomes. We compare survival models based on neural networks (CoxCC and CoxTime) and random survival forests (Interaction Forests). A Cox model, including an adaptive LASSO penalty, is used as a benchmark. Specific metrics for individualized treatment recommendations are used: the C-for-Benefit, the E50-for-Benefit, and RMSE for treatment benefit. We conduct an extensive simulation study using 2 different data generation processes incorporating nonlinearity and interactions up to the third order. The models are applied to gene expression and clinical data from 2 breast cancer studies. The machine learning-based methods show reasonable performances on the simulation data sets, especially in terms of discrimination for Interaction Forests and calibration for the neural networks. They can be used to evaluate individualized treatment effects from randomized trials when nonlinear and interaction effects are expected to be present.