Deep Survival Analysis With Latent Clustering and Contrastive Learning.

Area of Science:

• Biostatistics
• Machine Learning
• Data Science

Background:

• Survival analysis is crucial for time-to-event data but faces challenges with censored data.
• Current methods often overlook correlations between data instances, limiting predictive accuracy.
• Addressing censored data in survival analysis remains a significant research problem.

Purpose of the Study:

• To propose a novel Deep Survival Analysis model, DSACC, that integrates representation learning, latent clustering, and survival prediction.
• To enhance survival prediction by effectively utilizing correlations within data, particularly for censored instances.
• To improve the handling of censored data by leveraging information from uncensored samples within learned clusters.

Main Methods:

• Developed a unified framework for joint optimization of representation learning, latent clustering, and survival prediction.
• Introduced a novel contrastive loss function utilizing learned clusters to relate censored and uncensored data.
• Employed latent clustering to reveal and incorporate cluster distribution structures in the latent representation space.

Main Results:

• DSACC achieved advanced performance on four clinical datasets.
• Demonstrated superior C-index values (ranging from 0.6350 to 0.7943) compared to existing methods.
• Showcased improved Integrated Brier Score (IBS) values (ranging from 0.1120 to 0.2028), indicating better survival prediction accuracy.

Conclusions:

• The proposed DSACC model effectively addresses the challenge of censored data in survival analysis.
• Jointly optimizing representation learning and clustering significantly enhances survival prediction capabilities.
• DSACC offers a promising approach for improving the accuracy and reliability of survival analysis models, especially in clinical applications.