Predictive modeling of proliferative vitreoretinopathy using automated machine learning by ophthalmologists without coding experience

This retrospective cohort study (n=506 eyes) assessed the feasibility of ophthalmologists without coding experience designing machine learning (ML) algorithms to predict proliferative vitreoretinopathy (PVR) using automated ML (AutoML). Two ophthalmologists used a MATLAB application to build and evaluate ML algorithms. Models including pre-existing PVR as an input showed better area under the curve (AUC) for PVR prediction. A quadratic support vector machine (SVM) model (AUC 0.90, sensitivity 63.0%, specificity 97.8%) and an optimized Naïve Bayes algorithm (AUC 0.81, sensitivity 54.3%, specificity 92.4%) were developed. The study concludes that developing ML models for PVR prediction by non-coding ophthalmologists is feasible, though data scientist input may be needed to address class imbalance.

Fares Antaki, Ghofril Kahwati, Julia Sebag, Razek Georges Coussa, Anthony Fanous, Renaud Duval, Mikael Sebag