Abstract
Purpose :
To utilize pre-operative patient features obtained from electronic medical records to assess the accuracy of machine learning (ML) algorithms to predict trabeculectomy outcomes.
Methods :
Demographic, ocular and systemic health data resulting in 40 features from 220 consecutive trabeculectomy operations on 175 patients were input into four machine learning algorithms. The primary outcome classification was surgical failure at 1 year, defined as IOP > 21 or <5 at two consecutive visits after 3 months, less than a 20% IOP reduction, or a need for reoperation for glaucoma. Eyes that had not failed by the above criteria and were not receiving supplemental medical therapy to lower IOP were considered a complete success and the remaining eyes considered failures. Up-sampling was carried out to equalize the frequency of the underrepresented class. Random forest (RF), artificial neural network (ANN), logistic regression (LR) and support vector machine (SVM) algorithms were evaluated using accuracy, sensitivity, specificity and area under the curve (AUC) metrics with 10-fold nested cross-validation. The model with the best performance was further optimized using recursive feature elimination and hyper-parameter tuning. The data set was divided using 60% training/validation and 40% testing for evaluation of the final optimized model. The features most influencing the accuracy of each algorithm were identified using mean decrease in accuracy (MDA).
Results :
123 (55.90%) of the operations were classified as surgical failures and 97 (44.10%) as complete successes. The accuracy and AUC were highest in the RF algorithm, followed by SVM, LR and ANN as seen in Table 1. The accuracy and AUC of the final optimized RF model were 0.65 and 0.70, respectively. The most influential features on the accuracy of the final RF model are shown in Fig 1.
Conclusions :
To our knowledge, this is the first study to leverage the use of machine learning algorithms to predict trabeculectomy outcomes. The four ML algorithms performed similarly, with the RF algorithm providing moderately higher accuracy and AUC. Future work will focus on combining trabeculectomy outcomes from other institutions to increase sample size and provide more generalized predictive algorithms.
This is a 2021 ARVO Annual Meeting abstract.