# Deep Recurrent Survival Analysis

**Authors:** Kan Ren, Jiarui Qin, Lei Zheng, Zhengyu Yang, Weinan Zhang, Lin Qiu, Yong Yu  
**arXiv:** [1809.02403](https://arxiv.org/abs/1809.02403)  
**Date:** 2018-09-07  
**Summary page:** https://www.summarizepaper.com/en/paper/1809.02403v2/

## Key points

- Survival analysis is crucial for modeling time-to-event data and handling censorship in various fields such as clinical research and information systems.
- Existing methodologies for survival analysis often rely on counting-based statistics or assumptions about event probability distribution with respect to time, lacking consideration for sequential patterns within the feature space.
- Deep Recurrent Survival Analysis integrates deep learning techniques with survival analysis methods to predict conditional probabilities at a fine-grained level, addressing censorship issues and capturing time dependencies effectively.
- This model can predict the likelihood of true event occurrences and estimate survival rates over time without assuming specific forms of event probability distribution, fitting various sophisticated data distributions.
- Experimental evaluations have shown that the Deep Recurrent Survival Analysis model outperforms state-of-the-art solutions across multiple metrics in real-world tasks from different domains.

## Summary

, , , , 

In the realm of statistical research, survival analysis stands out as a crucial area for modeling time-to-event data while effectively handling censorship. This method has found widespread application in diverse fields such as clinical research and information systems, where survivorship bias is a common challenge. Over the years, various approaches have been developed for survival analysis, spanning from traditional statistical methods to more advanced machine learning models. However, existing methodologies often rely on counting-based statistics applied to segmented data or make assumptions about the event probability distribution with respect to time. Additionally, only a few works take into account sequential patterns within the feature space. In response to these limitations, a novel approach known as Deep Recurrent Survival Analysis has been proposed in this paper. The Deep Recurrent Survival Analysis model integrates deep learning techniques for predicting conditional probabilities at a fine-grained level of the data, coupled with survival analysis methods to address censorship issues. By capturing time dependencies through modeling the conditional probability of events for each sample, this model can predict the likelihood of true event occurrences and estimate survival rates over time – specifically, the probability of non-occurrence of events for censored data. One key advantage of this model is its ability to fit various sophisticated data distributions without assuming any specific form of event probability distribution. In experimental evaluations across three real-world tasks from different domains, the Deep Recurrent Survival Analysis model has demonstrated superior performance compared to state-of-the-art solutions across multiple metrics. Authored by Kan Ren, Jiarui Qin, Lei Zheng, Zhengyu Yang, Weinan Zhang, Lin Qiu, and Yong Yu; this research presents a significant advancement in survival analysis methodology and showcases its potential for enhancing predictive accuracy in time-to-event modeling scenarios.