Auditable unit-aware thresholds in symbolic regression via logistic-gated operators

Symbolic regression offers the promise of readable equations but often struggles to represent unit-aware thresholds that drive clinical decisions. We propose logistic-gated operators (LGOs)—differentiable gates with learnable location and steepness—embedded as typed primitives in a genetic-programming SR engine. Training is performed in standardized space; thresholds are mapped back to physical units for audit. On three public health datasets (MIMIC-IV ICU, eICU, NHANES), LGO hard identifies cut-points that appear clinically plausible: 4/13 fall within 10% of guideline anchors, 7/13 within 20%; the remaining gates tend to shift toward extreme-risk or early-warning regimes. On the derived high-risk classification tasks (based on score thresholds), compared with AutoScore, LGO achieves comparable or modestly improved area under the receiver operating characteristic curve (AUROC) and calibration while providing explicit thresholds rather than point tables. On smooth UCI benchmarks, gates are frequently pruned, helping to preserve parsimony. The approach yields relatively compact symbolic equations with auditable, unit-aware thresholds that may support clinical decision-making.