An Algorithm for Generating Individualized Treatment Decision Trees and Random Forests.

Area of Science:

• Biomedical Informatics
• Statistical Learning
• Clinical Decision Support

Background:

• Precision medicine aims to tailor treatments using patient data, but high-dimensional datasets challenge traditional methods.
• Existing approaches often struggle to identify complex interactions between treatments and numerous patient covariates.
• Novel methods are emerging to directly estimate individualized treatment rules (ITR) by optimizing clinical rewards.

Purpose of the Study:

• To propose a novel decision tree algorithm and reward function for directly maximizing clinical rewards in precision medicine.
• To develop an ensemble method, ITR random forests, to improve upon single decision tree rules.
• To create a soft probability-based decision rule that aids physician judgment.

Main Methods:

• Developed a new reward function and a decision tree algorithm for direct reward maximization.
• Implemented an ensemble approach using ITR random forests by averaging single decision trees.
• Utilized support vector machines (SVM) and decision trees as foundational methods for ITR estimation.

Main Results:

• The proposed ITR forest and tree methods were evaluated through simulations.
• Performance was assessed using a randomized controlled trial (RCT) dataset of 1385 diabetes patients.
• The methods were also applied to a large electronic medical record (EMR) cohort of 5177 diabetes patients.

Conclusions:

• The developed ITR forest and tree methods offer a promising approach for precision medicine.
• The soft probability output allows for flexible integration into clinical decision-making.
• These novel algorithms, implemented in R, can enhance the identification of optimal treatment strategies.