Planning an Artificial Intelligence-Based Diabetic Retinopathy Screening Program Using Human Centered Design

Abstract

Purpose : To describe the use of a human-centered design process when planning an artificial intelligence (AI)-based strategy for diabetic retinopathy (DR) screening using EyeScreen (EyeNuk, Inc, Woodland Hills, CA) in an academic health system.

Methods : We adapted the UK Design Council’s Double Diamond model to include five phases (Frame, Observe, Define, Build, and Evaluate; (Figure 1) in the design process. In the first Diamond, the Frame and Observe phases are used to explore the context as broadly as possible to gain understanding of the problem. In the Define phase, the problem is reframed based on an analysis of the information gained from the first two steps to inform the development of solutions. In the second Diamond, a wide array of possible solutions are considered, prioritized, and developed (Build). Evaluation is where those concepts in the Build phase are refined, iterated on, and chosen for implementation, and implementation follows after.

Results : Findings from the pre-implementation phases (Frame, Observe, Define, Build, Evaluate) are presented in this report. Over a 12-month period (July 1, 2021 to June 2022), there were over 1,500 patients with diabetes within a single Family Medicine (FM) practice and an estimated 30-40 patients per month who were overdue for DR screening at UI Health (Chicago, Illinois). Only 11% of referrals from FM to Ophthalmology for DR screening resulted in completed visits. While both FM and Ophthalmology reported a shared goal of preventing vision loss through screening and referral, the two groups viewed the DR screening itself very differently and had different definitions of success. For example, in FM, the screening would help them meet clinical target metrics for their patients while in the Department of Ophthalmology the utility of the screening was to improve patient adherence to visits. The results of the first three phases suggested the need to build optionality (FM provider visit or screening visit first), iterate workflows, and prioritize the referral of patients with a positive DR screen to Ophthalmology clinic.

Conclusions : An adapted Double-Diamond model can be utilized to guide the development of a stakeholder-supported AI-based DR screening program. Future work will discuss implementation, post-implementation program evaluation and modifications.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

 

Figure 2 demonstrates a proposed workflow for implementing EyeScreen at UI Health.