Predictive modeling of proliferative vitreoretinopathy using automated machine learning by ophthalmologists without
Fares Antaki1,2,3, Ghofril Kahwati4,5, Julia Sebag1
1Department of Ophthalmology, Université de Montréal, Montreal, QC, Canada.
Scientific Reports
|November 12, 2020
Summary
Ophthalmologists can now predict proliferative vitreoretinopathy (PVR) using machine learning (ML) without coding. Automated ML tools show feasibility, with models achieving high accuracy in predicting PVR after retinal detachment surgery.
Area of Science:
- Ophthalmology
- Medical Informatics
- Machine Learning
Background:
- Proliferative vitreoretinopathy (PVR) is a significant cause of vision loss after retinal detachment surgery.
- Accurate prediction of PVR is crucial for timely intervention and improved patient outcomes.
- Current methods for PVR prediction may not fully leverage the potential of machine learning.
Purpose of the Study:
- To assess the feasibility of ophthalmologists without coding experience developing machine learning (ML) algorithms to predict postoperative proliferative vitreoretinopathy (PVR).
- To evaluate the performance of ML models in predicting PVR using clinical data from electronic health records.
- To identify key clinical features associated with the development of PVR.
Main Methods:
- Retrospective cohort study of 506 eyes undergoing pars plana vitrectomy for rhegmatogenous retinal detachment (RRD).
- Two ophthalmologists utilized an interactive MATLAB application for automated machine learning (AutoML) model development.
- Clinical data from electronic health records were used, with univariate feature selection identifying relevant predictors.
Main Results:
- Machine learning models incorporating pre-existing PVR demonstrated superior predictive performance (AUC 0.90 for quadratic SVM).
- A quadratic SVM model achieved 97.8% specificity and 63.0% sensitivity in predicting postoperative PVR.
- An optimized Naïve Bayes model without pre-existing PVR achieved an AUC of 0.81, with 92.4% specificity and 54.3% sensitivity.
Conclusions:
- Developing ML models for PVR prediction is feasible for ophthalmologists lacking coding expertise.
- Automated ML tools can empower clinicians to build predictive models from electronic health records.
- Addressing class imbalance in real-world clinical data may require collaboration with data scientists for optimal ML model performance.
