Postoperative data leakage in AI-based IOL power prediction formulas: a hidden source of bias

This study addresses a critical bias in non-optical formulas purely based on AI models and designed to predict the implanted power using the postoperative spherical equivalent (SE) as an input. Two predictive models were developed using the XGBoost algorithm on a training set of 4588 eyes implanted with Finevision IOLs. One model (AI Pred. SE) was trained to predict the postoperative SE from the implanted power and biometric data, while the second model (AI Pred. Power) was designed to predict the implanted IOL power using the achieved SE and biometric data. Both models were evaluated on an independent testing set using standard performance metrics. The logic of the optical behavior was then assessed for each model, by varying the relevant input (SE for AI Pred. Power, Power for AI Pred. SE) across a broad range of values. On standard evaluation, the AI Pred. Power model showed the lowest prediction error (Friedman test p < 0.0001), outperforming conventional formula. However, despite its accuracy in predicting the implanted IOL power, this architecture returned a predicted power that was largely invariant and non-physiologically responsive to changes in the target refraction regardless of the target refraction entered as an input, for a given eye. This finding underscores the importance of excluding real postoperative refractive outcomes from any evaluation processes or formula design. We hypothesize that this common evaluation paradigm for pure-AI formulas creates a fundamental and undetectable bias, producing models that are optimized for retrospective prediction but are incapable of prospective clinical application.