ESCRS - PP22.01 - Development Of Prediction Models For Posterior Capsule Rupture With Machine Learning Based On The European Registry Of Quality Outcomes For Cataract And Refractive Surgery (Eurequo)

Development Of Prediction Models For Posterior Capsule Rupture With Machine Learning Based On The European Registry Of Quality Outcomes For Cataract And Refractive Surgery (Eurequo)

Published 2022 - 40th Congress of the ESCRS

Reference: PP22.01 | Type: Free paper | DOI: 10.82333/6c3s-n998

Authors: Maartje Segers* 1 , Ron Triepels 2 , Paul Rosen 3 , Rudy Nuijts 1 , Frank van den Biggelaar 1 , Ype Henry 4 , Ulf Stenevi 5 , Marie-José Tassignon 6 , David Young 7 , Anders Behndig 8 , Mats Lundström 9 , Mor Dickman 1

1University Eye Clinic,Maastricht University Medical Center+,Maastricht,Netherlands, 2Department of Data Analytics and Digitalisation,Maastricht University,Maastricht,Netherlands, 3Department of Ophthalmology,Oxford Eye Hospital,Oxford,United Kingdom, 4Department of Ophthalmology,Amsterdam UMC,Amsterdam,Netherlands, 5Department of Ophthalmology,Sahlgrenska University Hospital,Göteborg,Sweden, 6Department of Ophthalmology,Antwerp University Hospital,Antwerp,Belgium, 7Department of Mathematics and Statistics,University of Strathclyde,Glasgow,United Kingdom, 8Department of Clinical Sciences, Ophthalmology,Umea University,Umea,Sweden, 9Department of Clinical Sciences, Ophthalmology,Lund University,Lund,Sweden

Purpose

To evaluate the performance of different probabilistic classifiers to predict posterior capsule rupture (PCR).

Setting

Clinics affiliated with the European Registry of Quality Outcomes for Cataract and Refractive Surgery (EUREQUO).

Methods

Three probabilistic classifiers were constructed to predict the probability of PCR, including a Bayesian Network (BN), a Logistic Regression (LR) model, and a Multi-Layer Perceptron (MLP) network. The classifiers were trained on a sample of 2,853,376 surgeries reported to the European Registry of Quality Outcomes for Cataract and Refractive Surgery (EUREQUO) between 2008 and 2018. The performance of the models was evaluated by nested cross-validation based on the Area Under the Precision-Recall Curve (AUPRC). Direct risk factors of PCR were identified by inspecting the Markov Blanket of PCR in the BN.

Results

The MLP network predicted PCR overall the best (AUPRC 13.1 ± 0.41%), followed by the BN (AUPRC 8.05 ± 0.39%), and the LR model (AUPRC 7.31 ± 0.15%). Direct risk factors for PCR include preoperative best corrected visual acuity, year of surgery, operation type, target refraction, anesthesia, white cataract, corneal opacities, and other ocular comorbidities.

Conclusions

Our results indicate that the MLP network predicts PCR the best. Although the precision is relatively low at high recall, the network appears to perform better than existing scoring models in the literature. Consequently, it is expected that implementating the MLP network in clinical practice will decrease PCR rate even further than existing scoring models. Future research should verify this in a clinical setting.