Deep Learning Model Optimizes Treatments for Two Survival Outcomes
New adaptive method maximizes joint survival probability using deep neural networks.
In a new paper published on arXiv, researchers Kun Ren, Yifan Cui, and Wen Su tackle a complex problem in personalized medicine: determining the best treatment for each patient when there are two correlated survival outcomes (bivariate survival). For example, a cancer therapy might aim to maximize both progression-free survival and overall survival. The team introduces a deep neural network framework that learns optimal individualized treatment rules (ITRs) to maximize the joint survival probability beyond two fixed time points, $(t_1, t_2)$, while accounting for right censoring common in clinical trials.
Their method models treatment rules as stochastic policies, coupling marginal accelerated failure time (AFT) models via a link function (e.g., a copula) to capture the dependence between the two survival endpoints. To enhance robustness and efficiency, they propose an adaptive prediction-powered learning approach that leverages auxiliary predictions from other machine learning models. This allows the method to correct for potential biases and handle limited sample sizes. The work is rooted in statistical machine learning (stat.ML) and reinforces practical decision-making from randomized trial data, offering a path toward truly individualized therapy recommendations.
- Handles bivariate survival outcomes (e.g., progression-free and overall survival) with right censoring via deep neural networks.
- Uses stochastic policies and marginal accelerated failure time models linked by a copula to capture dependence between endpoints.
- Introduces an adaptive prediction-powered method to improve robustness by leveraging auxiliary ML predictions.
Why It Matters
Enables more personalized treatment recommendations from trials by optimizing for multiple correlated survival outcomes simultaneously.