The Loss or the Architecture? What Matters in Survival Analysis with Competing Risks
A 23-dataset benchmark finds cumulative-incidence loss, not backbone choice, drives competing-risks performance.
The authors separate architecture from training loss for competing-risks survival analysis, testing three losses on four backbones across 20 tabular and three pathology-image datasets. Modeling cumulative incidence directly under censoring weights wins in both modalities and contributes about twice as much as choosing a better backbone on competing-risk data. Proper scoring rules reveal the gap, while the concordance index does not; RealMLP or TabM trained with the cumulative-incidence loss achieves the best Brier score.
- Three losses tested on four backbones across 23 datasets
- Cumulative-incidence loss outweighs backbone choice for competing risks
- Tabular networks lead on tables; vision encoders lead on images
- C-index does not separate models, while Brier score does
- RealMLP or TabM with incidence loss posts the best Brier score
From the original article
1 min read at arxiv.orgOften, \emph{eg} in healthcare and predictive maintenance, what matters is which of several competing causes will strike first, and with what probability: survival analysis with competing risks. Corresponding deep learning methods are developed as bundles: a new architecture together with a new training loss, tested on one to five datasets against a handful of baselines.
By Julie Alberge · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.