A regression tree approach to identifying subgroups with differential treatment effects

Area of Science:

• Biostatistics
• Clinical Trial Design
• Machine Learning in Medicine

Background:

• Discovering effective treatments for complex diseases like cancer is challenging.
• Identifying patient subgroups with enhanced treatment response is crucial for regulatory approval.

Purpose of the Study:

• Introduce novel regression tree algorithms for subgroup identification.
• Address limitations of existing methods in handling complex clinical trial data.

Main Methods:

• Extend the generalized unbiased interaction detection and estimation (GUIDE) approach.
• Incorporate treatment as a linear predictor, chi-squared tests, and Poisson regression for proportional hazards modeling.
• Handle censored data and missing predictor variables.

Main Results:

• Developed three new regression tree algorithms practically free of selection bias.
• Algorithms are applicable to randomized trials with multiple treatments and complex data.
• Generated importance scores and confidence intervals for treatment effects.

Conclusions:

• The new algorithms offer a robust approach for identifying patient subgroups with differential treatment effects.
• These methods enhance the analysis of clinical trial data, aiding in personalized medicine strategies.
• The techniques are validated using both real and simulated datasets.