Identification of interactions of binary variables associated with survival time using survivalFS

Area of Science:

• Biostatistics
• Computational Biology
• Genomics

Background:

• Medical studies frequently analyze time-to-event data, such as survival or relapse times.
• Identifying interactions between binary variables is crucial for accurate prediction of disease recurrence.
• Exhaustive interaction testing is often infeasible, necessitating advanced analytical procedures.

Purpose of the Study:

• To present survivalFS, an ensemble method based on logic regression, to address model instability.
• To introduce methods for quantifying the importance of interactions and individual variables in time-to-event models.
• To evaluate the performance of survivalFS and its importance measures against existing methods like random survival forests.

Main Methods:

• Developed an ensemble method, survivalFS, utilizing logic regression to handle complex interactions.
• Introduced novel measures for quantifying interaction and variable importance, considering multivariate data structures.
• Proposed an adapted version of Harrel's concordance index for performance evaluation.
• Conducted simulation studies and applied the method to urinary bladder cancer genotype data.

Main Results:

• survivalFS effectively identifies interactions associated with time to an event.
• The proposed importance measures provide robust quantification of variable and interaction significance.
• survivalFS demonstrated superior performance compared to random survival forests in the analyzed datasets.
• The method successfully identified relevant interactions in urinary bladder cancer genotype data.

Conclusions:

• survivalFS offers a stable and effective approach for analyzing time-to-event data with complex interactions.
• The developed importance measures enhance the interpretability of logic regression models in survival analysis.
• survivalFS represents a valuable advancement for identifying predictive factors in medical research, particularly in genomics.