OR WAIT null SECS
Adding AI technology using deep-learning systems in a semi-automated model to the process has been shown to be equally effective as the assessment performed by trained human graders for detecting DR.
Diabetic retinopathy (DR) is a vision-stealing complication of diabetes that has been effectively assessed by human graders of fundus photographs using tele-ophthalmology screening.
Adding artificial intelligence (AI) technology using deep learning systems in a semi-automated model to the process has been shown to be equally effective as the assessment performed by trained human graders for detecting DR. The added bonus is that the technology is substantially more cost-effective.
The study investigators explained that “before the deep learning era, a few UK studies suggested that automated DR grading using feature-based learning might be cost-effective as a DR screening tool. Since the advent of deep learning, a few studies have shown that [it] could have comparable diagnostic performance to human graders in detecting DR.”
The investigators from four institutions in Singapore and one in Guangzhou, China, took that step forward to test that hypothesis by conducting a cost-minimization analysis to evaluate the potential savings of AI deep learning approaches compared with the costs associated with human analysis.
The study, which was performed in Singapore, looked at two deep learning approaches, i.e., one semi-automated and the other fully automated. The former served as a triage filter before secondary human assessment of fundus photographs and the latter did not include human assessment, they explained.
Their analysis included 39,006 consecutive patients with diabetes in a national diabetic screening program.
The investigators compared the actual cost of screening the patient cohort by human graders with the simulated cost of the two automated models. The parameters used in the model were the DR prevalence rates, DR screening costs in each screening model, cost of medical consultation, and the diagnostic performance to determine the sensitivity and specificity. The primary outcome of the study was the total cost of each model.
The investigators reported, “Our study showed that the semi-automated model is the least expensive of the three models, followed by the fully automated model and then the human assessment.” The investigators published their results in The Lancet (https://doi.org/10.1016/S2589-7500(20)30060-1).
Specifically, they found that the cost associated with the semi-automated screening program was $62US per patient annually compared with $66US per patient annually for the fully automated program and $77US per patient annually for the human assessment model.
“The savings to the Singapore health system associated with switching to the semi-automated model are estimated to be $489,000, which is roughly 20% of the current annual screening cost,” they reported.
The projected savings by 2050 is $15 million considering that by that time point there would be 1 million people with diabetes in Singapore.
Corresponding author: firstname.lastname@example.org
The authors are from the Singapore National Eye Centre, National University of Singapore, Duke-NUS Medical School, Tan Tock Seng Hospital, all in Singapore; and the Sun Yet-Sen University, Guangzhou, China.