Deep Learning for Survival Analysis: A Review

TL;DR

This review systematically catalogs deep learning approaches to time-to-event analysis, framing the problem through established SA targets such as $S_T(t)$, $h_T(t)$, and $H_T(t)$. It analyzes 61 DL-based survival methods, classifying them by model class, loss, and network architecture, and maps them onto data modalities and survival task types, including competing risks and multi-state settings. The authors highlight that most methods focus on single-risk, right-censored data and adapt DL techniques from CV/NLP, with limited coverage of interval censoring and more complex outcomes, while noting reproducibility and benchmark challenges. An open-source, interactive table is provided to foster community contributions, and the paper offers guidance on interpretability, evaluation, and future directions such as diffusion-based approaches for survival modeling.

Abstract

The influx of deep learning (DL) techniques into the field of survival analysis in recent years has led to substantial methodological progress; for instance, learning from unstructured or high-dimensional data such as images, text or omics data. In this work, we conduct a comprehensive systematic review of DL-based methods for time-to-event analysis, characterizing them according to both survival- and DL-related attributes. In summary, the reviewed methods often address only a small subset of tasks relevant to time-to-event data - e.g., single-risk right-censored data - and neglect to incorporate more complex settings. Our findings are summarized in an editable, open-source, interactive table: https://survival-org.github.io/DL4Survival. As this research area is advancing rapidly, we encourage community contribution in order to keep this database up to date.

Figures (7)

• Figure 1: PRISMA diagram for literature screening of deep learning-based survival methods.
• Figure 2: Schematic neural architecture for competing risks in survival analysis using shared and cause-specific subnetworks. $\psi()$ transforms the model output (e.g., hazard rate) to the final outcome (e.g., cumulative incidence functions (CIFs)).
• Figure 3: Schematic neural architecture for multimodal data input in survival analysis using separate subnetworks for all modalities. Their outputs are reshaped and concatenated to align dimensions. $\psi()$ transforms the model output to the final outcome. The X-ray scan is obtained from irvin2019chexpert.
• Figure 4: Absolute frequencies of model classes among all 61 methods reviewed.
• Figure 5: Absolute frequencies of neural network architectures among all 61 methods reviewed.