KAPLAN-HR: Kolmogorov-Arnold Network for Survival Analysis
There's this innovative method called KAPLAN-HR that uses Kolmogorov-Arnold Networks, or KANs, for estimating hazards in survival analysis without relying on specific parameters. It looks at how different factors and time affect the hazard using B-splines. The simpler version of this model produces generalized additive models (GAMs), while more advanced setups can capture complex interactions and changes over time. One cool thing is that the accuracy of the model depends only on how smooth the KAN is, which helps avoid issues with high-dimensional data. In tests on six clinical datasets, KAPLAN-HR proved to be a strong alternative to older methods like the Cox model, which requires manual setup for interactions.
Key facts
- KAPLAN-HR is a B-spline Kolmogorov-Arnold Network for nonparametric hazard estimation.
- It models the conditional hazard as a joint function of covariates and time.
- A single-layer KAPLAN-HR recovers a generalized additive model.
- Deeper architectures capture interactions and time-varying effects.
- The convergence rate depends only on smoothness of the KAN representation, not covariate dimension.
- It mitigates the curse of dimensionality for KAN-representable targets.
- Evaluated on six clinical benchmark datasets.
- Classical methods like the Cox model require manual specification of interactions.
Entities
—