Heterogeneous domain adaptation survival analysis with partially observed outcomes via dictionary learning and distribution alignment.

Abstract

Domain adaptation (DA) aligns learned representations so that predictive models generalize across domains with differing population characteristics, particularly when limited target-domain data prevents training domain-specific models. DA is especially relevant in healthcare, where patient populations and measurement protocols vary across hospitals, phenotypic groups, and disease stages. Clinical risk prediction often involves time-to-event data, where outcome events such as death require long follow-up and are frequently affected by dropout or loss to follow-up, resulting in partially observed survival outcomes. However, DA methods for survival analysis under heterogeneous feature spaces and partial supervision has received limited attention. This study proposes a Heterogeneous Domain Adaptation (HDA) framework for survival analysis with partially observed outcomes. By integrating dictionary learning, Maximum Mean Discrepancy (MMD), and the Cox Proportional Hazards model, the framework learns a shared latent representation that aligns source and target feature distributions despite differences in their feature dimensionalities, thereby improving survival discrimination in the target domain. Across both simulation experiments and real-world case studies, HDA-CoxPH achieved absolute C-index improvements ranging from 0.027 to 0.130 compared to the top-performing benchmark models. The proposed framework consistently outperformed traditional survival models while maintaining competitive predictive performance with up to 50% unlabeled target data. Overall, our study showed that the proposed HDA framework can enhance survival prediction in data-scarce settings with partially observed outcomes by leveraging heterogeneous data sources through dictionary learning and MMD-based distribution alignment. The implementation of our framework is publicly available at: https://github.com/dannilin2601/hda_survival_analysis.