Abstract
Purpose :
With teleretinal imaging (TRI) emerging as a viable screening tool, current generalized screening guidelines for diabetic retinopathy (DR) may not be appropriate for all patients with diabetes mellitus (DM). A mathematical modeling-based study was conducted to derive personalized screening recommendations.
Methods :
A partially observable Markov decision process (POMDP) model, a common technique used for sequential decision making over a patient’s lifetime with uncertain disease progression, was created to determine personalized screening recommendations for patients with DM. For each hypothetical patient, the model makes recommendations of either no action (NA), TRI, or clinical screening (CS) on a yearly basis, with the objective of optimizing a weighted combination of accumulated QALYs and cost. If a patient is deemed to have sight-threatening DR after TRI, an immediate CS occurs. The model was applied to a group of hypothetical 40-year-old African American male patients. Sensitivity analysis was conducted to understand the impact of different levels of compliance, personal trade-off preference between QALYs and cost, and sensitivity of TRI on personalized recommendations.
Results :
During the lifetime of patients with 100% compliance rate, TRI was recommended 89.82% (95% CI=89.28%-90.35%) of the time, NA was recommended 10.18% (95% CI=9.65%-10.72%) of the time, and CS was not recommended at all. When the sensitivity of TRI for diagnosing sight-threatening disease was increased to 90% from 74.87%, the recommendation for TRI increased by 0.09%. For patients with 50% compliance, the recommendation for CS increased by 7.95%, TRI decreased by 6.79%, and NA decreased by 1.16% (Figs1 and 2).
Conclusions :
The POMDP model indicates a strong preference for TRI-based screening versus CS for patients with 100% compliance. Patients’ individual compliance appears to be the driving factor affecting recommended screening type. As non-compliance increases, the recommendation for the more accurate CS increases to offset the risk of blindness. This model can be utilized as a decision support tool to provide a personalized screening strategy that caters to each patient's unique needs and preferences.
This is a 2020 ARVO Annual Meeting abstract.