Neural Network-Based Prediction of Post-Operative Visual Outcomes Following Secondary Pediatric Intraocular Lens Implantation

TL;DR

A machine learning model was developed to predict visual outcomes in children after intraocular lens implantation, showing promising results for clinical decision-making.

Contribution

The novel contribution is a proof-of-concept neural network model for predicting post-operative visual outcomes in pediatric intraocular lens implantation.

Findings

The model achieved 88.2% accuracy, 88.9% sensitivity, and 87.5% specificity in predicting visual outcomes.

ROC curve analysis showed AUC values ranging from 0.885 to 0.942 across training, validation, and test sets.

The model demonstrates feasibility despite limited dataset diversity, supporting future personalized strategies in pediatric cataract care.

Abstract

What are the main findings?
A proof-of-concept neural network model was developed to predict visual outcomes after secondary intraocular lens (IOL) implantation in children with congenital cataracts.The model demonstrated encouraging predictive performance across training, validation, and test sets, suggesting feasibility despite the limited dataset.

A proof-of-concept neural network model was developed to predict visual outcomes after secondary intraocular lens (IOL) implantation in children with congenital cataracts.

The model demonstrated encouraging predictive performance across training, validation, and test sets, suggesting feasibility despite the limited dataset.

What is the implication of the main finding?
This work underscores the potential of machine learning to support clinical decision-making for secondary IOL implantation, an area currently lacking predictive tools.Broader, multi-center datasets and models restricted to preoperative variables will be essential to validate and translate this approach into clinical practice.

This work underscores the potential of machine learning to support clinical decision-making for secondary IOL implantation, an area currently lacking predictive tools.

Broader, multi-center datasets and models restricted to preoperative variables will be essential to validate and translate this approach into clinical practice.

Objectives: To develop a proof-of-concept machine learning (ML) neural network model to predict post-operative visual outcomes in children with congenital cataracts undergoing intraocular lens (IOL) implantation, thereby guiding the optimal timing for IOL insertion. Determining the ideal timing and predicting outcomes for IOL implantation in children remains clinically complex due to variability in eye development and measurement accuracy. Methods: Retrospective analysis using a publicly available dataset from 110 children diagnosed with congenital cataracts, who underwent IOL implantation at the Eye and ENT Hospital of Fudan University. A neural network model with a hidden layer of 10 nodes was developed in MATLAB 2024a using the scaled conjugate gradient algorithm. Input variables included demographic and clinical features; the target was achieving visual acuity greater than 20/40. Performance metrics were evaluated using cross-entropy loss, sensitivity, specificity, and accuracy. Results: Training completed after 14 epochs with the test set reaching the highest performance metrics: 88.2% accuracy, 88.9% sensitivity, and 87.5% specificity. ROC curve analysis showed AUC values of 0.942 (training), 0.920 (validation), 0.885 (test), and 0.917 (overall). Conclusions: The neural network effectively predicted post-operative visual outcomes, offering potential clinical utility in guiding IOL implantation decisions. Despite limitations in dataset diversity, this study lays the foundation for future development of personalized strategies in pediatric cataract care.

Figures