KULLBACK-LEIBLER-BASED DISCRETE FAILURE TIME MODELS FOR INTEGRATION OF PUBLISHED PREDICTION MODELS WITH NEW...

Area of Science:

• Biostatistics
• Health Informatics
• Survival Analysis

Background:

• Time-to-event data prediction is challenged by rare events, small sample sizes, and high dimensionality.
• External prediction models can enhance internal prognosis prediction but often assume data similarity, which is frequently not the case.
• Existing integration methods face limitations due to data heterogeneity, sharing, and privacy constraints.

Purpose of the Study:

• To propose a novel failure time integration procedure for combining external prediction models with internal data.
• To address challenges of data heterogeneity, sharing, and privacy in model integration.
• To improve the performance of prognosis prediction using disparate data sources.

Main Methods:

• Developed a discrete hazard-based Kullback-Leibler discriminatory information measure to quantify discrepancies between external models and internal datasets.
• Proposed a failure time integration procedure utilizing this discrepancy measure.
• Validated the method through asymptotic properties analysis and simulation studies.

Main Results:

• The proposed integration method demonstrated superior performance compared to methods relying solely on internal data.
• Simulation results confirmed the advantage of the new approach in improving prediction accuracy.
• The method successfully enhanced prediction performance on a kidney transplant dataset by integrating local data with a national registry model.

Conclusions:

• The novel failure time integration procedure effectively improves prognosis prediction by leveraging external models, even with heterogeneous data.
• The Kullback-Leibler discrepancy measure provides a robust way to account for differences between external and internal data sources.
• This approach offers a valuable tool for enhancing predictive modeling in healthcare, particularly for rare events and limited internal data.