Purchase this article with an account.
Gongyu Zhang, Zeyu Guan, Edward Korot, Reena Chopra, Gabriella Moraes, Siegfried Wagner, Livia Faes, Dun Jack Fu, Hagar Khalid, daniel Araujo ferraz, Konstantinos Balaskas, Daniel Alexander, Pearse Andrew Keane; Optimising Treatment of Neovascular Age-related Macular Degeneration using Reinforcement Learning. Invest. Ophthalmol. Vis. Sci. 2020;61(7):1628.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Develop a reinforcement learning algorithm to optimise treatment decisions of neovascularAMD. The algorithm provides two-dimensional treatment suggestions including whether the patient should be injected and when to schedule the next follow up visit.
We used a previously developed deep learning OCT segmentation model as input for our model. Segmented tissues included as input were subretinal fluid (SRF), intraretinal fluid (IRF). The model was optimized to reach a goal of zero volume of IRF and SRF for three consecutive visits. Patient admissions were modelled by Markov decision processes. Action space, problem state, reward function was developed in this framework. Q function of temporal difference learning was updated using iteration and sampling from the historical patient data. The output is whether the patient should be injected and when to schedule the next followup visit.
The mean, [median] number of treatment visits for the AI policy to obtain 3 consecutive visits with 0 IRF and SRF was 16,  , which was less than clinicians’ 25, . Therefore, AI policy can help patients reduce visit time.Based on historical data, when deciding not to inject, the majority of clinicians in our dataset scheduled a followup visit between 40 to 45 days. For comparison, when deciding not to inject, the majority of the time, our reinforcement learning AI model scheduled a followup visit between 90 to 95 days.When the decision was made to inject, the clinician and AI model followup times were 45 to 50 days, and70 to 75 days respectively.
A data-driven approach using reinforcement learning was developed to optimize the treatment of neovascular age-related macular degeneration. The algorithm demonstrates the generation of treatment courses for AMD patients to reduce total IRF and SRF with less visits and longer inter-visit intervals than clinicians from historical data.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only