Researchers' SVM survival trees map leukemia risk with sharp geographic boundaries
New two-stage method separates spatial hotspots from clinical covariates without model assumptions.
Drew Lazar and Aye Aye Maung (arXiv:2608.13847) present a nonparametric two-stage method that separates spatial risk from patients' clinical characteristics. In the first stage, a clinical survival tree fit to the covariates alone generates leaf Nelson-Aalen cumulative hazard residuals, placing censored survival data on a clinically adjusted scale without imposing a functional form on the clinical hazard. A second tree, fit to these residuals on the coordinates, then recovers the spatial structure. Both stages use kernel dipole-splitting survival trees, so the resulting spatial risk map is piecewise constant with sharp, possibly curved boundaries. On the LeukSurv leukemia data, the method agrees with a Bayesian Gaussian random field frailty about where risk is elevated while resolving sharp adjacencies the smooth surface averages away, and an unadjusted spatial analysis misattributes clinical variation to location. Simulations with known zones, including graded violations of exogeneity, locate the point at which the spatial and clinical contributions cease to be separately identifiable, with the smooth benchmark degrading in parallel as that point is approached.
- Two-stage nonparametric method separates spatial from clinical risk using kernel dipole-splitting SVM survival trees
- Produces piecewise-constant risk maps with sharp curved boundaries, tested on LeukSurv leukemia data
- Simulations sweep graded exogeneity violations to locate identifiability limits, with smooth benchmarks degrading in parallel
Why It Matters
Enables accurate geographic disease hotspot detection without confounding by patient clinical profiles, improving public health resource allocation.