Abstract
Purpose :
Among the treatment options available for DME, intravitreal injections of anti-VEGFs have been the most practiced option. However, response to the anti-VEGF injection varies from patient to patient. In order to assist clinicians in optimizing treatment, we developed an automatic algorithm that predicts the potential change in CST for a given DME patient before administrating the anti-VEGF injection.
Methods :
For the development and verification of the machine learning (ML) models, we collected data from a total of 906 DME patients who received anti-VEGF treatment. Each patient’s multiple visit data points were collected where both pre- and post-injection OCT scans were acquired using CIRRUS™ HD-OCT 4000/5000 (ZEISS, Dublin, CA) with signal strength ≥ 0.4. The patients’ data were divided into train and test sets containing 775 and 131 patients’ data respectively.
Four benchmark models including Linear Regression, Support Vector Regression (SVR) with Radial Basis Function (RBF) kernel, SVR with linear kernel and Random Forest (RF) were trained using clinical parameters such as patient’s age, gender, near BCVA value, distance vision (DV) BCVA value, pre-injection CST value, the other eight ETDRS grid values, number of previous anti-VEGF injections, and number of days after injection. Fig 1 shows the schematic diagram of the ML model development.
Using the test set, the performance of the CST change prediction algorithms was evaluated in terms of correlation coefficient (CC). The performance of all 4 models was further evaluated for four different scenarios - number of injections administered: ≥ 1, ≥ 2, ≥ 3 and > 3.
Results :
SVR with linear kernel model achieved best performance in terms of CC with 0.65, 0.73, 0.75 and 0.85 for four scenarios respectively. Fig 2 shows the performance of all four ML models for the four scenarios. Linear regression, RF and SVR with RBF kernel model perform in the descending order of correlation coefficient.
Conclusions :
In this study, we demonstrated how various ML regression models perform, with best CC of 0.85 with linear SVR for > 3 injections, in predicting the change in CST value for DME patients undergoing anti-VEGF treatment. This may help specialists to better plan the treatment of DME patients.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.