level: research
randomized trials with multiple treatments often face challenges when outcomes are bivariate survival times. this paper presents a method to derive optimal individualized treatment rules that maximize the joint survival probability beyond two fixed time points. the approach uses deep neural networks and models treatment rules as stochastic policies. it couples marginal accelerated failure time models through a link function to capture dependence between the two survival outcomes, while accounting for right censoring.
to improve robustness, the authors introduce an adaptive prediction-powered learning technique. this method leverages auxiliary predictions from machine learning models to enhance decision making. the adaptive component adjusts how much the auxiliary predictions are used based on their reliability, aiming to reduce variance and bias in the estimated treatment rules. the framework is designed for settings where some labeled data is available along with a larger set of unlabeled data with predicted outcomes.
the proposed method is evaluated through simulations and a real data application. results show that the adaptive prediction-powered approach outperforms standard methods that ignore auxiliary predictions or use them naively. the deep neural network policy class allows flexible modeling of complex treatment assignment mechanisms. the method provides a practical tool for personalized decision making in clinical trials where two survival endpoints are of interest.
why it matters: this method helps data scientists build more reliable personalized treatment strategies when two time-to-event outcomes must be balanced, using auxiliary predictions to improve efficiency.