Predicting OCT-morphological anti-VEGF treatment response in patients with neovascular age-related degeneration using artificial intelligence

Abstract

Purpose : To make individualized predictions of patients’ morphological response to intravitreal injections with Bevacizumab, Ranibizumab and Aflibercept using a deep learning algorithm. For this purpose, we determined the minimal amount of data points necessary to predict the response type of a patient and the best performing classification method.

Methods : Training of a machine learning algorithm using different models (random forest algorithm, support vector machine, Logistic Regression, Ridge regression, fully connected neural network) to predict treatment response in treatment naïve patients after 90 days. Patients were categorized as good responder, poor responder and non-responder based on their resolution of fluids segmented by a previously built deep U-net based semantic segmentation ensemble algorithm. For the different algorithm models we used 413 patients as a training, 116 patients as a validation and 222 patients as a test set. All patients were treatment naïve.

Results : Using only the baseline OCT as input, F1 scores were best for the random forest algorithm model in the good and poor responder group (F1=0.9, precision=0.47, recall=0.9). Non responders were more difficult to predict albeit still achieving F1=0.43 using the support vector machine training model. Best overall results were achieved using both the baseline oct and the oct after the first intravitreal injection with the random forest algorithm achieving an F1 score of 0.9 for the good/poor responder group and 0.48 for the non-responder group. As expected, the predictions got more accurate when more visits were used as input data.

Conclusions : The predictive algorithm shows highly efficient prediction of patients’ treatment response especially in the good and poor responder group, while non-response prediction was moderate. Non-response is generally not very common whence accurate predicting will require a lot more input data in the future. This model can be used to further train and differentiate which medication is best suited for the present morphological structure to move towards a more personalized treatment regimen.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Performance of treatment response prediction using different machine learning models.