Out-Of-Distribution Detection For Enhanced Safety And Optimization In Machine Learning -Based Icl Sizing

To determine if out-of-distribution (OOD) detection can identify preoperative cases where machine-learning (ML) postoperative vault predictions for implantable collamer lenses (ICL) are likely to be unreliable, supporting safer clinical decision-making.

Private clinical practice

Two independent ocular biometry cohorts were analyzed (Dataset A: 152 eyes, Antwerp, Belgium; Dataset B: 140 eyes, London, UK). A Ridge Regression (RR) model was trained on Dataset A to predict postoperative vault. Simultaneously a One-Class Support Vector Machine (OCSVM) was trained on the same preoperative feature space to quantify Out-of-Distribution (OOD) deviation. Performance was validated on a held-out internal test set (Test-A: 37 eyes) and the external UK cohort (Dataset B). Prediction error was measured as Mean Absolute Error (MAE) and absolute error (AE) per eye. Eyes were stratified into "In-Distribution" (ID) and "OOD-flagged" based on a predefined OCSVM threshold, with MAE compared between strata using non-parametric testing.

The RR model trained on the Antwerp data (Dataset A) generalized well, with no significant difference in error distributions between the internal test set and the London cohort (MAE Test-A: 156 ± 116 µm vs MAE Dataset B: 198 ± 174 µm; p > 0.32).  The OCSVM stratified eyes by prediction risk with OOD scores positively correlated with per-eye AE in Test-A (Pearson r = 0.11) and dataset B (r = 0.25 ).  When stratified by the OOD flag across a mixed sample of both cohorts, the OOD-flagged group (n=77) exhibited a 60% higher mean error compared to the In-Distribution group (n=100) (240 ± 196 µm vs 150 ± 122 µm, respectively; p < 0.05).

OOD detection using OCSVM provides a case-level warning signal that aligns with vault prediction error, including in a geographically distinct external cohort. This supports a practical safety layer for ML-based ICL sizing. Alongside a vault estimate, clinicians could preoperatively receive an interpretable reliability flag indicating when the model is operating outside its training population and predictions should be interpreted with caution. Beyond safety, this approach could be adapted to match patients with the specific ICL sizing formula that best fits their biometric profile, leading to more optimized, individualized selection.