Optimizing Treatment Decision Estimation for Right-Censored Survival Data Through Parameter Transfer Learning

Accurately estimating treatment effects is crucial for designing optimal treatment plans in personalized medicine, especially in the presence of right-censored survival data. We propose a parameter transfer learning method for estimating treatment effects on right-censored survival data, which leverages multi-source auxiliary data to enhance the prediction accuracy and robustness of the target model. This method constructs multiple source models by extracting shared parameters from other datasets and uses a smoothed concordance index function specifically designed for right-censored survival data to estimate candidate model parameters. To enhance performance, a leave-one-out cross-validation criterion is applied to optimize model averaging weights. Theoretically, we have demonstrated that under mild conditions, the proposed method asymptotically achieves the highest smoothed concordance index when the target model is misspecified, and ensures model weight consistency when the target model is correctly specified. Simulation studies confirm the advantages of our proposed method in reducing bias and enhancing prediction accuracy, particularly with right-censored and heterogeneous data. Its application to the SUPPORT (Study to Understand Prognoses and Preferences for Outcomes and Risks of Treatments) extension dataset, support2, further demonstrates its strong potential in personalized clinical decision-making.