Model selection for survival individualized treatment rules using the jackknife estimator.

Area of Science:

• Biostatistics
• Machine Learning in Medicine
• Precision Medicine

Background:

• Precision medicine aims to personalize treatments based on individual patient data to maximize clinical outcomes.
• Identifying individualized treatment rules (ITRs) is key, especially for time-to-event outcomes where handling censored data is critical.

Purpose of the Study:

• To propose a novel jackknife estimator for the value function to accommodate right-censored survival data in binary treatment scenarios.
• To enable the estimation of optimal ITRs using machine learning methods by addressing censoring through inverse probability of censoring weighting (IPCW).

Main Methods:

• Developed a jackknife estimator incorporating IPCW adjustment for right-censored time-to-event data.
• Estimated optimal ITRs using Random Survival Forest (RSF) and Cox proportional hazards (COX) models.
• Compared RSF and COX-derived ITRs against a zero-order model (ZOM) using a Z-test and simulation studies.

Main Results:

• COX outperformed RSF in smaller sample sizes; RSF performance improved with larger sample sizes, particularly when relationships were non-linear.
• The proposed estimator demonstrated normal distribution across various simulation scenarios and censoring rates.
• Application to non-small cell lung cancer data indicated the ZOM performed best, suggesting no need for treatment tailoring in this specific dataset.

Conclusions:

• The jackknife approach effectively estimates the value function with right-censored data, performing well under small to moderate censoring.
• Winsorizing estimated survival weights for IPCW calculation enhances estimator stability.
• The findings underscore the importance of robust methods for ITR estimation in survival analysis and highlight potential scenarios where simpler models suffice.