Gaussian Adaptive Thick‑Lens Iol Navigation & Error Landscaping (Gatinel): Patient‑Specific Prediction Intervals And Improved Refractive Accuracy In A Large Hold‑Out Dataset

To present the principles and clinical performance of Gatinel (Gaussian Adaptive Thick‑lens IOL Navigation & Error Landscaping), an adaptive IOL power calculation framework designed to remain robust when IOL information is incomplete (known/unknown IOL model; A‑constant available/unavailable), and to provide patient‑specific prediction intervals for individualized risk‑aware planning.

Rothschild Foundation Hospital

Gatinel combines a thick‑lens vergence model with an adaptive effective-lens-position (ELP) strategy. Using a Gaussian uncertainty framework for each candidate IOL power, the postoperative spherical equivalent (SE) is estimated, along with uncertainty derived from sensitivity‑based error propagation and heteroscedastic model terms. This enables Error Landscaping, i.e., mapping the predicted error distribution across IOL powers to compute a patient‑specific prediction interval and the probability of meeting the refractive target. Performance was evaluated using mean error (ME), SD, MAE, and the proportion within ±0.25 D / ±0.50 D / ±1.00 D. 

In an independent hold‑out dataset (N=5204 eyes), after ME centering, Gatinel achieved SD 0.487 D, MAE 0.341 D, 48.5% within ±0.25 D, 78.2% within ±0.50 D, and 96.3% within ±1.00 D. Comparator formulas showed lower accuracy: Haigis (SD 0.634 D; MAE 0.449 D; 68.4% within ±0.50 D), SRK/T (SD 0.654 D; MAE 0.470 D; 65.4%), Holladay 1 (SD 0.631 D; MAE 0.455 D; 66.5%), and Hoffer Q (SD 0.613 D; MAE 0.447 D; 66.6%). Gatinel improved the proportion within ±0.50 D by +9.8 to +12.8 percentage points compared with classic formulas.

Gatinel improved refractive accuracy compared with widely used formulas on a large hold‑out dataset and provides patient‑specific prediction intervals through a Gaussian uncertainty framework and error‑landscaping approach, supporting individualized counseling and risk‑aware IOL power selection beyond point prediction alone.