Adaptive Transfer Learning for Time-to-Event Modeling with Applications in Disease Risk Assessment

Area of Science:

• Biostatistics
• Machine Learning
• Health Informatics

Background:

• Modeling time-to-event outcomes in small-sample settings presents significant challenges.
• Existing methods struggle with data heterogeneity and potential shifts between source and target datasets.

Purpose of the Study:

• To propose CoxTL, a novel transfer learning approach for time-to-event analysis in small-sample settings.
• To enhance predictive accuracy by accounting for covariate and concept shifts.

Main Methods:

• CoxTL is based on the Cox proportional hazards model, incorporating density ratio and importance weighting.
• It addresses multi-level data heterogeneity, including covariate and coefficient shifts.
• The method is robust to potential model misspecification.

Main Results:

• In simulations, CoxTL showed higher predictive accuracy, especially with multi-level heterogeneity.
• When predicting End-Stage Renal Disease (ESRD) risk in Hispanic populations, CoxTL improved the C-index by up to 6.76% compared to target-only models.
• CoxTL outperformed state-of-the-art transfer learning methods by up to 17.94% in C-index.

Conclusions:

• CoxTL effectively utilizes source data to improve time-to-event predictions in small target populations.
• The method's robustness to data heterogeneity makes it suitable for real-world applications where traditional Cox models falter.