Machine learning and neural networks successfully predict glaucoma surgery outcomes

Machine learning (ML) and neural networks (NN) have been successfully used to predict outcomes of glaucoma surgery, according to a poster presented at AAO 2023.

The study, conducted over a 9-year period, utilized structured data from electronic health records (EHR) of 1445 patients who underwent incisional glaucoma surgery at Stanford between 2013 and 2022.

The input features for the models included demographic information, prior diagnoses and procedure codes, medications, and findings from eye examinations. Criteria for defining a successful surgical outcome included the absence of follow-up glaucoma surgery and intraocular pressure (IOP) levels between 6 and 19 without an increase in glaucoma medications.

Out of the 2172 surgeries analyzed, 1320 were deemed unsuccessful. The study found that the best ML model for predicting overall surgical failure was the random forest model, with an accuracy of 71% and an area under the curve (AUC) of 66%. The best performing NN model achieved an accuracy of 69% with an AUC of 67%.

The research also highlighted that predicting IOP-related failure outcomes (AUC 70%) proved more straightforward compared to predicting medication outcomes (AUC 63%) and the need for additional surgery (AUC 59%).

Reference
Barry S, et al. Predicting Glaucoma Surgery Outcomes Using AI. Presented at: AAO 2023.