Using artificial intelligence to predict the risk for posterior capsule opacification after phacoemulsification

Abstract

Purpose: To apply artificial intelligence models to predict the occurrence of posterior capsule opacification (PCO) after phacoemulsification.

Setting: Farabi Eye Hospital, Tehran, Iran.

Design: Clinical-based cross-sectional study.

Methods: The posterior capsule status of eyes operated on for age-related cataract and the need for laser capsulotomy were determined. After a literature review, data polishing, and expert consultation, 10 input variables were selected. The QUEST algorithm was used to develop a decision tree. Three back-propagation artificial neural networks were constructed with 4, 20, and 40 neurons in 2 hidden layers and trained with the same transfer functions (log-sigmoid and linear transfer) and training protocol with randomly selected eyes. They were then tested on the remaining eyes and the networks compared for their performance. Performance indices were used to compare resultant models with the results of logistic regression analysis.

Results: The models were trained using 282 randomly selected eyes and then tested using 70 eyes. Laser capsulotomy for clinically significant PCO was indicated or had been performed 2 years postoperatively in 40 eyes. A sample decision tree was produced with accuracy of 50% (likelihood ratio 0.8). The best artificial neural network, which showed 87% accuracy and a positive likelihood ratio of 8, was achieved with 40 neurons. The area under the receiver-operating-characteristic curve was 0.71. In comparison, logistic regression reached accuracy of 80%; however, the likelihood ratio was not measurable because the sensitivity was zero.

Conclusion: A prototype artificial neural network was developed that predicted posterior capsule status (requiring capsulotomy) with reasonable accuracy.

Financial disclosure: No author has a financial or proprietary interest in any material or method mentioned.