Selecting combinations of visual function measures to identify severity of diabetic retinopathy

Abstract

Purpose : To establish which combinations of up to 3 visual function tests have the best discriminative performance for classifying diabetic retinopathy (DR) severity stage.

Methods : In this cross-sectional study, participants underwent 9 visual function tests. Diabetes mellitus (DM) status was determined using self-report, medication history and HbA1c measurements. Fundus, ultra-widefield and OCT images were graded for DR and diabetic macular edema (DME). Eyes were classified into four groups: healthy with no DM (n = 324); DM no DR (n=116); DR no DME (n=91); DR with DME (n=52).

Three classification tasks were set:
A) distinguishing DM no DR from no DM,
B) DR no DME from DM no DR,
C) DR with DME from DR no DME.

Ensemble machine learning models were fitted for all 1-, 2- and 3-way combinations of visual function variables (129 models for each task). Models also contained age and sex. Performance was assessed with area under the receiver operating characteristics curve (AUC).

Results : Participants had median age 62, 64% were female. Model performance varied substantially among combinations, with AUC ranging from 0.60 to 0.92 for task A, 0.50 to 1.00 for task B and 0.50 to 0.93 for task C.

For task A, high performing models contained distance visual acuity (all 10 of the top 10 ranked models contained this measurement) and 5/10 contained low-luminance distance visual acuity. For task B mesopic microperimetry was in 8/10 models. For task C, distance visual acuity (7/10) and Smith-Kettlewell low luminance near visual acuity (6/10) featured prominently. High ranking models all combined two or three measurements. Single measurement models were ranked 40th or lower for each task.

Conclusions : Models combining information from 2 or more visual function measurements achieved the highest performance when discriminating DR severity stages, outperforming single measurement models. The most useful visual function measurements differed depending on severity stage. For detecting early-stage disease, distance visual acuity was important. For detecting retinopathy among those with DM, mesopic microperimetry ranked highly and for detecting DME among those with DR, distance visual acuity and Smith-Kettlewell low luminance near visual acuity performed well. This study gives evidence to support the selection and combination of visual function tests for use in both clinical and trial contexts.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.