Risk controlled decision trees and random forests for precision Medicine.

Area of Science:

• Biostatistics
• Machine Learning
• Clinical Decision Support

Background:

• Individualized treatment rules (ITRs) aim to maximize patient benefit but may overlook risks like hypoglycemia in type 2 diabetes (T2D).
• Current ITR methods often prioritize benefit maximization, potentially leading to unacceptable risk levels for certain patient subgroups.

Purpose of the Study:

• To develop and evaluate novel statistical methods for identifying risk-controlled individualized treatment rules (rcITRs).
• To ensure that maximized treatment benefits do not exceed a predefined risk threshold.

Main Methods:

• A penalized recursive partitioning algorithm was developed to create risk-controlled decision trees (rcDTs).
• Risk-controlled random forests (rcRFs) were proposed as an extension for enhanced modeling robustness.
• Three variable importance measures were introduced to aid clinical decision-making.

Main Results:

• Simulation studies confirmed the robustness and effectiveness of the proposed rcRF modeling approach.
• The methods were successfully applied to analyze data from the DURABLE diabetes trial.
• An R package is available for implementing the rcDT and rcRF procedures.

Conclusions:

• The proposed rcITR methods, including rcDT and rcRF, provide a framework for optimizing treatment strategies while controlling patient risk.
• These methods are applicable to both randomized controlled trials and observational studies, enhancing clinical decision support.
• The developed tools facilitate the identification of safer and more effective ITRs across various medical applications.