Discriminative optimization erases biological relational structure in radiomic feature selection: A multi-objective...

Radiomic features are not independent: they occupy a dependency graph whose communities track acquisition sequence and tumour anatomy. Sparse selection discards that graph, but the size of the loss, whether it depends on the model discriminating, and whether it can be avoided are unknown.

We define four normalised structural-loss functionals - edge-weight, spectral, community, topological - measuring how much of a feature-dependency graph a selected subset preserves. Graphs are estimated within each training fold by neighbourhood selection. Nine selectors, including four graph-aware and multi-objective comparators, are evaluated on three endpoints spanning a wide signal range (MGMT methylation, n = 574; IDH mutation, n = 501; WHO grade, n = 501; 1158 IBSI features), with external validation in two cohorts (n = 432) and replication across 45 datasets covering 23 pathologies and three modalities. Operating points are chosen on inner validation splits.

Sparse selectors lose more than 75 % of graph edge weight in 42 of 45 datasets. The loss is unrelated to discriminative success - 0.905 where AUC ≥ 0.75 versus 0.865 below 0.65 (p = 0.43) - and persists under label permutation. Laplacian-regularised selection preserves substantially more structure at indistinguishable AUC (IDH: 0.717 versus 0.874, ΔAUC = -0.001, p = 0.90) and occupies 57.8 % of the joint Pareto front. The four functionals are empirically redundant (ρ ≥ 0.97 in 45 of 45 datasets).

Structural erosion is intrinsic to sparsity-inducing optimisation, not a symptom of weak signal, and graph-Laplacian regularisation avoids most of it at no cost in performance.