TabSurv: Adapting Modern Tabular Neural Networks to Survival Analysis

Problem
Survival analysis on tabular data is a well-established field, yet existing deep learning methodologies are often tailored to specific tasks, which hinders the transferability of techniques across domains. This paper addresses the gap in adapting modern tabular neural network architectures for survival analysis, particularly focusing on the limitations of current approaches that may restrict performance and generalizability. The work is presented as a preprint and has not undergone peer review.

Method
The authors introduce TabSurv, a framework that leverages contemporary tabular neural network architectures for survival analysis. TabSurv employs either the Weibull distribution or a non-parametric approach for survival prediction. A key innovation is the SurvHL, a novel histogram loss function designed to handle censored data effectively. The architecture includes a baseline feed-forward network and deep ensembles of multi-layer perceptrons (MLPs). Notably, the ensemble components are trained in parallel, optimizing survival distribution parameters prior to averaging, which enhances the diversity of predictions across the ensemble. This method contrasts with traditional ensemble techniques that may not prioritize parameter optimization before aggregation.

Results
TabSurv was empirically evaluated across 10 diverse real-world survival datasets. The results indicate that TabSurv consistently outperforms established classical and deep learning baselines, including Random Survival Forests (RSF), DeepSurv, DeepHit, and SurvTRACE. Specifically, deep ensembles utilizing Weibull parametrization achieved the highest average rank based on the concordance index (C-index), demonstrating superior predictive performance. The paper provides detailed performance metrics, showcasing significant effect sizes that underscore the efficacy of the proposed method compared to the aforementioned baselines.

Limitations
The authors acknowledge several limitations, including the potential overfitting of deep ensembles due to their complexity, particularly in smaller datasets. They also note that while the Weibull distribution provides a robust framework, it may not capture all survival data characteristics, suggesting that further exploration of alternative distributions could be beneficial. Additionally, the reliance on parallel training may introduce computational overhead, which could be a concern in resource-constrained environments. The paper does not address the scalability of the method to extremely large datasets or its performance in high-dimensional feature spaces.

Why it matters
The implications of this work are significant for the field of survival analysis, as it demonstrates how modern tabular neural networks can be effectively adapted for this purpose. By providing a robust framework that integrates advanced loss functions and ensemble techniques, TabSurv enhances the predictive capabilities of survival models. This advancement opens avenues for further research into the application of deep learning in survival analysis, potentially leading to improved methodologies that can be generalized across various domains. The public availability of the TabSurv implementation facilitates further exploration and validation by the research community.

Authors: Stanislav Kirpichenko, Andrei Konstantinov, Lev Utkin
Source: arXiv:2605.03944
URL: https://arxiv.org/abs/2605.03944v1