Ponente
Descripción
This study investigates the capability of a Multi-Layer Perceptron (MLP) neural network (NN), initially trained to predict intraocular lens (IOL) power for a specific lens model in cataract surgery, to generalize its predictions to other IOL models. A retrospective analysis was conducted using biometrical data, including axial length (AL) and average corneal power (Kav), as well as predicted refraction, postoperative stable refraction, and implanted IOL power from patients who underwent uneventful cataract surgery with in-the-bag implantation of IOLs from five different models. Four experiments were designed to evaluate the NN's predictive performance: (1) training and testing the NN on a single IOL model; (2) training on one model and testing on another; (3) training the NN on a dataset comprising 500 cases from each model and testing on 500 cases of a single model; and (4) training the NN on a database containing 20% of cases from each model and testing on the complete dataset for each model. The study included 26,936 eyes (one per patient). Statistical analysis revealed non-normal data distribution at the 0.01 significance level, with AL values drawn from different distributions, while K values were not. Results demonstrated that NNs accurately predicted IOL power for the model they were trained on. In some cases, they could also predict power for other models. By subtracting a prediction bias, any NN could predict outcomes for all models. In conclusion, a NN trained on one IOL model can predict the power of another model, though in certain cases, prior data may be required to determine and correct for bias.
