Combining techniques for screening and evaluating interaction terms on high-dimensional time-to-event data.

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• High-dimensional molecular data, such as from microarrays, is crucial for predicting patient survival.
• Existing multivariate risk prediction models integrate parameter estimation and variable selection.
• Incorporating interaction terms into these models presents a significant computational challenge.

Purpose of the Study:

• To develop and evaluate methods for incorporating interaction terms in high-dimensional time-to-event prediction models.
• To address the computational expense of evaluating all possible interactions.

Main Methods:

• Utilized a boosting technique for effect estimation.
• Employed building blocks for interaction pre-selection: resampling, random forests, and orthogonalization.
• Conducted simulation studies and analyzed real-world data.

Main Results:

• The combined strategy effectively detected true main effects and interactions across various scenarios.
• Identified challenges in detecting interactions not represented by main effects, but showed promising results.
• Real-world data indicated that interaction effect sizes may not always be sufficient to enhance prediction performance despite biological relevance.

Conclusions:

• Screening interactions using random forests is a feasible and valuable approach for identifying relevant two-way interactions.
• Resampling, random forests, and orthogonalization significantly enhance interaction pre-selection.
• Defined the limits of interaction detection based on necessary effect sizes, highlighting the need to leverage existing and new methods.