Segmented Versus Global Optimization of Intraocular Lens Constants Across Axial Length: A Multi-Lens, Multi-Formula...

Background/Objectives: Intraocular lens (IOL) constants are conventionally optimized globally across an entire calibration dataset, assuming that systematic prediction error is independent of axial length (AL). This assumption is known to fail in short and long eyes. We evaluated whether AL-segmented constant optimization reduces refractive prediction error, whether the benefit depends on formula family, and whether changepoints discovered in one population transfer to a disjoint one. Methods: In a pooled 6451-eye training dataset (6 IOL models), the corrected Akaike Information Criterion selected the optimal number (0-4) and positions of AL changepoints for six formulas (classic three-constant Haigis, new-Haigis H1/H2, SRK/T, Hoffer Q, Holladay 1), enforcing ≥40 eyes and ≥2.0 mm per segment. Locked changepoints were applied to two disjoint single-lens test subgroups (Vivinex, n = 887; SA60AT, n = 821; three-center retrospective cohort) and independently re-discovered within each. Root-mean-square prediction error (RMSE) reduction was assessed by bootstrap confidence interval, Diebold-Mariano test, and Wilcoxon signed-rank test. Results: Segmentation reduced training RMSE for all six formulas, particularly for Hoffer Q (-3.9%) and Holladay 1 (-2.8%; confidence intervals excluding zero, Diebold-Mariano p < 0.0001). With locked changepoints, both formulas again showed the largest test-dataset improvements (3.0-6.3%; confidence intervals excluding zero in all four lens-workflow combinations), followed by SRK/T (0.9-1.3%, up to 2.5% with independently discovered changepoints). Haigis-family formulas showed smaller reductions (0.2-1.6%) with confidence intervals including zero in several combinations. Conclusions: AL-segmented optimization benefits Hoffer Q and Holladay 1 most robustly-formulas lacking a direct anterior chamber depth predictor-with a smaller benefit for SRK/T and an inconsistent benefit for Haigis-family formulas. Changepoints from a pooled training population transfer usefully, though not optimally, to individual lens models.