Selecting measures of visual function to classify diabetic retinopathy status: a cross-sectional study

To identify combinations of up to three visual function tests with the best performance for classifying diabetic retinopathy (DR) severity stage. To describe in detail the measurements from a comprehensive set of visual function tests.

1901 eyes (1032 participants) underwent nine visual function tests. Fundus, ultra-widefield and optical coherence tomography images were graded for DR and diabetic macular oedema (DMO). Three classification tasks were set: (1) distinguishing diabetes mellitus (DM) no DR from healthy with no DM, (2) DR no DMO from DM no DR and (3) DR with DMO from DR no DMO. Ensemble machine learning models for all one-way, two-way and three-way combinations of visual function variables were compared using area under the curve (AUC).

The top 30 models for each task achieved high accuracy, with AUC ≥0.94. For task 1, 17/30 top models contained distance visual acuity. Pelli-Robson contrast sensitivity and low luminance visual acuity also featured highly. For task 2, 19/30 models contained mesopic microperimetry. Near visual acuity, matrix microperimetry and reading index featured highly. For task 3, 17/30 models contained distance visual acuity. Smith-Kettlewell low luminance near visual acuity and near visual acuity featured highly. In a subset of eyes where perimetry was not performed, reading index featured in 22, 21 and 22 of the top models for tasks 1, 2 and 3 respectively.

These findings will enable researchers and those planning clinical trials to select the best combination of visual function tests for distinguishing stages of diabetic eye disease.