Small-Sample Survival Modeling via Transfer Learning and Representation Calibration

Abstract

Survival analysis is a critical statistical methodology utilized to model the time until a specific event of interest occurs. In recent years, deep learning has revolutionized survival modeling by relaxing linear assumptions and capturing complex non-linear variable interactions. However, the efficacy of deep neural networks in survival analysis is heavily contingent upon the availability of large-scale datasets. In numerous real-world applications, such as rare disease prognosis and novel industrial component failure testing, gathering sufficient event data is prohibitively expensive or practically impossible, leading to a severe small-sample problem. Under such constraints, standard deep learning models are prone to severe overfitting and poor generalization. To address this fundamental limitation, this paper introduces a novel framework that integrates transfer learning with representation calibration. By leveraging extensive data from related source domains, our approach pre-trains a generic survival representation space. To counteract the inevitable domain shift between the source and target datasets, we propose a representation calibration mechanism that aligns the feature distributions without destroying the temporal dependencies inherent to survival data. Comprehensive experiments conducted on both clinical and industrial datasets demonstrate that our method significantly outperforms traditional survival models and standard fine-tuning approaches in low-resource settings. The calibrated transfer learning strategy improves prognostic accuracy and model robustness, offering a viable pathway for deploying advanced predictive analytics in domains constrained by minimal data availability