Flexible Transfer Learning in Deep Cox Models

ICLR 2026 Conference Submission21648 Authors

Keywords: Transfer learning; Survival analysis; Deep Learning; Cox model

TL;DR: We propose a flexible transfer learning framework for analyzing time-to-event data using Cox model.

Abstract: Prognosis prediction is an important topic in survival analysis. Historically, research aimed at predicting survival outcomes has largely been confined to individual datasets. These datasets often have limitations, such as rare event rates, small sample sizes, high dimensionality, and low signal-to-noise ratios. To overcome these limitations, integrated survival analysis and transfer learning has been proposed to improve prediction accuracy by incorporating external prediction models into the analysis of newly collected data. However, traditional integrated approaches, such as the integrated Cox proportional hazards model, often face limitations in prognostic prediction capabilities due to their dependence on the linearity and proportional hazards assumptions. In reality, the relationship between event times and risk factors can be intricate, often involving non-linear effects, influences that vary over time, and interactions. To effectively capture the complexities of integrated time-to-event data, it is essential to employ computationally efficient deep learning techniques.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 21648