Reproduce “CAPITAL: Optimal Subgroup Identification via Constrained Policy Tree Search” on OpenResearch

About this paper

Hengrui Cai, Wenbin Lu, Rachel Marceau West, Devan V. Mehrotra, Lingkang Huang

Researchers from the University of California Irvine, North Carolina State University, and Merck & Co., Inc. developed CAPITAL, a Constrained Policy Tree Search Algorithm, to identify optimal patient subgroups that maximize size while ensuring a pre-specified average treatment benefit. This method generates interpretable decision rules, demonstrating improved subgroup identification and adherence to clinical thresholds across simulations and real-world clinical trial data for continuous and survival outcomes.

Key insights

• The theoretical optimal subgroup selection rule identifies patients whose individual treatment effect exceeds a specific cut-point, determined by the target average treatment effect.
• Translating population-level constraints into individual rewards allows for effective policy learning using tree-based methods, ensuring both optimality and interpretability.
• A trade-off exists between maximizing subgroup size and imposing additional constraints, such as ensuring positive individual treatment effects, which can be managed by penalty parameters.

Summary by alphaXiv.