Artificial Intelligence Can Predict Retreatment Intervals in nAMD

Hrvoje Bogunovic, PhD, Department of Ophthalmology, Medical University of Vienna

Hrvoje Bogunovic, PhD

An automated artificial intelligence (AI)-based system was approximately 70% accurate at predicting optimal retreatment intervals for anti-VEGF therapy in patients with neovascular age-related macular degeneration (nAMD), according to findings presented at the 2018 ARVO Annual Meeting.

The approach, which utilized machine learning and optical coherence tomography (OCT) images, proved to be 68% sensitive and 73% specific for determining whether patients should receive a long (8 to 12 weeks) versus a short (4 to 6 weeks) retreatment interval. Once further validated on larger dataset, utilization of the AI-based system could lead to a personalized treatment plan for patients with nAMD.

"We developed a predictive model, which is using very accurate and precise biomarker identification, and then we applied the use of AI to allow us to enhance our clinical capabilities," lead investigator Hrvoje Bogunovic, PhD, Department of Ophthalmology, Medical University of Vienna, said during a presentation of the findings at the ARVO meeting. "For patients and clinicians right now, there is a lot of uncertainty about how the treatment is going to go. This could help, and provide prognostic value at the individual level.”

A predictive approach in nAMD is desperately needed, as treatment burden often leads to under-treatment with anti-VEGF therapy. Recent studies have examined new approaches to extend the retreatment interval, with strategies that include once every 12 weeks dosing and a treat-and-extend (T&E) strategy; however, the lingering question facing these approaches is optimal patient selection.

For the AI study, the first step was to train the machine, which was completed using data from the TREND study. In this 12-month phase 3b study, ranibizumab was administered in a T&E dosing strategy or in standard monthly doses. Disease activity assessments were conducted using OCT on intraretinal fluid (IRF) or subretinal fluid (SRF). The maximum treatment interview was 12 weeks.

The T&E strategy in the TREND trial was found to be non-inferior to monthly dosing. The best corrected visual acuity (BCVA) was 6.2 letters in the T&E arm and 8.1 letters for the monthly doses. Overall, 8.7 injections of ranibizumab were required in the T&E group versus 11.1 in the monthly arm.

When looking across patient groups, 18% of patients in the T&E arm required weekly dosing whereas 12% were able have treatment once every 12 weeks, with a great amount of variety for the remaining patients. Overall, 49% of patients had a short retreatment interval of 4 to 6 weeks and the remainder could receive treatment every 8 to 12 weeks.

The wide variety in the T&E arm of the TREND study, Bogunovic noted, raised the question of whether markers could be found to predict which patients were best suited for an every-12-week dose. For this, the researchers identified visual imaging biomarkers based on OCT. These included intraretinal layers, IRF, SRF, pigment epithelium detachment (PED), and hyper-reflective foci (HRF).

The OCT images were then introduced to a convolutional neural network, which analyzed the samples and matched specific anatomical characteristics in the images to the ability to receive a longer retreatment interval. The automated system could identify, localize, and quantify IRF and SRF. Additionally, using image registration, it could align intra and inter patient variability, for reproducibility.

To make the system more exact, over 300 spatio-temporal OCT imaging features were introduced to the machine learning algorithm. These were coupled with BCVA at baseline and month 1 along with age and gender.

Following 10-fold cross-validation, maximal retreatment interval could be predicted within 4.7 weeks using the AI system, with 95% agreement with the true interval seen in the TREND study. The most predictive biomarker for retreatment interval was found to be SRF in central 3 mm at 1 month. "The more subretinal fluid there was, the shorter the retreatment interval," noted Bogunovic.

Although these early findings are promising, and with more samples, the preciseness could increase, the effectiveness of this AI training set was limited by the short duration of the TREND study, which ran for 12 months, Bogunovic noted. Datasets with longer follow-up duration would be required to help predict the stability of the predictions. Additionally, the fact that retreatment guidelines also utilized SRF may have overstate the importance of this marker.

Further analyses will be required before the new software is able to reach the clinic.

Sign up to get frontline clinical insights directly to your inbox.

Related Coverage >>>

Artificial Intelligence Effectively Assesses Cell Therapy Functionality

Algorithm Detects Diabetic Retinopathy in Retinal Images with 97% Accuracy

Paolo Silva, MD: Deep Learning Algorithm Predicts Future DR Severity