Tree-based identification of subgroups for time-varying covariate survival data

Area of Science:

• Biostatistics
• Epidemiology
• Machine Learning

Background:

• Classification and regression trees (CART) identify sample subsets differing on outcomes.
• Standard CART models struggle with confounders and time-varying covariates in survival data.
• Existing CART variations address survival, time-varying covariates, or confounders individually, but not all together.

Purpose of the Study:

• To propose a novel method integrating CART with survival analysis, time-varying covariates, and confounder adjustment.
• To identify subsets of time-varying risk factors affecting survival outcomes.
• To demonstrate the technique's utility on real-world clinical trial data.

Main Methods:

• Developed a novel classification and regression tree approach.
• Incorporated adjustments for potential confounders.
• Accounted for time-varying covariates within right-censored survival data analysis.
• Applied the method to the Bypass Angioplasty Revascularization Investigation 2 Diabetes trial data.

Main Results:

• Successfully identified specific subsets of time-varying risk factors associated with time-to-event outcomes.
• Demonstrated the ability to adjust for confounders in complex survival data.
• Highlighted combinations of modifiable cardiac risk factors (e.g., smoking, blood pressure, HbA1c) linked to clinical outcomes.

Conclusions:

• The proposed method effectively identifies risk factor subsets impacting survival in complex epidemiological datasets.
• This approach enhances the analysis of time-varying covariates and confounders in survival studies.
• Offers a valuable tool for understanding modifiable risk factors in clinical outcomes.