A risk-stratified dynamic screening framework for age-related macular degeneration via a clustering-based partially observable markov decision process

Age-related macular degeneration (AMD) is the leading cause of irreversible visual impairment in elderly people. The multifactorial and often asymptomatic natures of AMD in the early stages, coupled with the high costs of screening, play a crucial role in guiding research studies toward finding proper screening policies. This paper presents a risk-stratified dynamic screening policy that considers the current personal characteristics and medical history of each individual. It utilizes a clustering-based framework with partially observable Markov decision process (POMDP) models to diagnose AMD disease in the early stage. The innovative part of the proposed framework is that the cluster assigned to each individual might be modified because of changes in his/her risk factors during decision-making process. In addition, a new simulator has been developed in which some performance criteria are obtained and different policies are appropriately compared. The observation-based screening schedule is implemented via a real ophthalmic dataset. The risk-stratified policy could more effectively increase quality-adjusted life years (QALYs) for individuals at high risk while decreasing unnecessary screening tests for those at low risk. The findings verify that the optimal policy obtained through the proposed framework outperforms both no-screening and annual-screening policies. Furthermore, the clustering-based framework exhibits minor sensitivity to the natural course of disease progression in healthy individuals and is completely robust against variations in the likelihood of false negative outcomes.