CD-Surv: a contrastive-based model for dynamic survival analysis.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Health Services Research

Background:

• Survival analysis is crucial for health service management, often utilizing longitudinal electronic health record (EHR) data.
• Existing survival analysis methods often use static data, ignoring visit correlations and latent patient trajectory representations, thus limiting performance.

Purpose of the Study:

• To propose an end-to-end contrastive-based model, CD-Surv, for dynamic survival analysis.
• To improve the understanding of patient treatment trajectories and dynamically predict survival probability.

Main Methods:

• Developed CD-Surv, an end-to-end contrastive learning model for survival analysis.
• Implemented two data augmentation strategies: mask generation and shuffle generation, to enhance real EHR treatment trajectories.
• Utilized contrastive learning between augmented and real trajectories to improve hidden representations.

Main Results:

• CD-Surv demonstrated superior performance compared to state-of-the-art baseline models.
• The model achieved improved evaluation metrics on two real-world datasets.
• Contrastive learning effectively enhanced the representation of patient treatment trajectories.

Conclusions:

• CD-Surv offers a powerful approach for dynamic survival analysis using longitudinal EHR data.
• The proposed method effectively captures latent representations of patient trajectories, leading to improved survival prediction.
• Contrastive learning with data augmentation is a promising strategy for enhancing survival analysis models.