Improving Preoperative Prediction of PCR

Posterior capsule rupture (PCR) occurs in approximately 1% of cataract surgery cases and is a feared event be­cause of its potential sight-threatening consequences. 

Now, work is underway leveraging data reported to the European Registry of Quality Outcomes for Cataract and Refractive Surgery (EUREQUO) to see if a machine learning model could improve preoperative identification of patients at increased risk of PCR.¹ 

Research conducted so far indicates such artificial intelli­gence-based classifiers outperform existing scoring systems and, therefore, if implemented in the clinic, may reduce the number of cases of PCR, said Rudy MMA Nuijts MD, PhD.

“Identification of patients at high risk for PCR is important so that these difficult cases can be assigned to experienced operators and for providing better communication to patients about their surgical risk,” said Dr Nuijts. “Existing scoring systems for PCR risk assessment are based on a subjective weighting of known risk factors. We were interested in seeing if data-driven risk assessment would estimate the probability of PCR more reliably and objectively.”

Three different probabilistic classifiers were constructed to estimate the probability of PCR before surgery. They included two discriminative models—a logistic regression model and a multi-layer perceptron network—and a gener­ative model known as a Bayesian network, which has the advantage that it can deal with missing data, Dr Nuijts said.

“Analysing data for PCR risk prediction is challenging because PCR is rare, there is an exponential number of feature combinations, and, as in all registries, we have missing data.”

All the classifiers were trained using data from about 2.8 million cases of cataract surgery performed between January 2008 and De­cember 2018. Performance of the classifiers was determined by calcu­lating area under the precision-re­call curve (AUPRC). The results showed the multi-layer perceptron network performed best followed by the Bayesian network and then the logistic regression model. The mean AUPRC for the three classifiers was 13.10, 8.05, and 7.31, respectively. 

“The multi-layer perceptron performs better than subjec­tive grading systems published in the literature,” Dr Nuijts said. “In addition, compared to a random classifier that has no knowledge of the patient or procedure, the machine learning classifiers are about 7 to more than 12 times better in generat­ing the probability.”

Analysis of an independence graph generated by the Bayesian network model was also performed to characterise direct and indirect risk factors for PCR. Direct risk factors included preoperative best-corrected visual acuity, year of surgery, operation type, anaesthe­sia, target refraction, other ocular comorbidities, white cataract, and corneal opacities.

Going forward, further work aims to validate and compare the machine learning models using data from other registries, such as the American Academy of Oph­thalmology’s Intelligent Research in Sight (IRIS) regis­try. Once the model is validated, Dr Nuijts envisioned the ideal situation as relevant information, including measurements from diagnostic systems, would auto­matically be entered into the electronic medical record, where they would be available to be processed in the probabilistic classifier.

“This would be the most efficient and beneficial way for surgeons to have an idea of the risk of PCR for any individual patient,” Dr Nuijts said.

Dr Nuijts spoke at the ESCRS Symposium during AAO Refractive Surgery Subspecialty Day 2023 in San Francisco, US. He acknowledged his co-authors—Dr Ron Triepels, Dr Maartje Segers, and Dr Mor Dickman.

1. Triepels RJMA, Segers MHM, Rosen P, et al. “Development of machine learning models to predict posterior capsule rup­ture based on the EUREQUO registry,” Acta Ophthalmol. 2023; 101(6): 644–650.

Rudy MMA Nuijts MD, PhD is Professor of Ophthalmology, Vice-Chair­man, and Director of the Cornea Clinic and the Center for Refractive Surgery at the University Eye Clinic, Maastricht, Netherlands. rudy. nuijts@mumc.nl

Ron JMA Triepels PhD is assistant professor at the Department of Data Analytics and Digitalization of Maastricht University, Maastricht, Nether­lands. r.triepels@maastrichtuniversity.nl

Maartje HM Segers MD is a PhD candidate at the University Eye Clinic, Maastricht University Medical Center+, Maastricht, Netherlands. maartje. segers@mumc.nl

Mor M Dickman MD, PhD is associate professor at the University Eye Clinic, Maastricht, Netherlands, and clinical co-director of the ESCRS registries. mor.dickman@mumc.nl