BACKGROUND AND AIMS In 2021, following evidence from several studies, the Diabetic Eye Service in Scotland introduced a recommendation of 2-year screening interval (2YSI) for patients whose most recent two screens had no diabetic eye disease (ND). Previously the maximum recommended screening interval was 1 year. The study aims were to assess the effectiveness and safety of the 2YSI policy and to compare effectiveness and safety between automatically generated and manual decisions for 2YSI.
MATERIALS AND METHODS: This study used retrospective data containing DR grades (as in table) and fundus images with screening dates from October 2005 to February 2017. Patient inclusion criteria were attendance at the DR screening programme in Scotland during this time. At least yearly screening was offered to people resident in Scotland with age >12 years with diabetes. The null hypotheses were: there is no difference in the distributions of the patient-level DR grades incident during 2YSI whether simulated using manual or automated detection. 50% of subjects were used to train a ResNet101 deep neural network architecture as an automated DR grader. Its outputs were input to a thresholded logistic regression model trained with 12.5% of subjects to detect ND. In the remaining 37.5% of subjects, the test set, we determined which screens would have been omitted if the 2YSI policy had been in place. The screen after two consecutive screens with ND was classed as omitted. Manual and the automated assessments of ND were used for this and so generated two sets of omitted screens. Pearson’s chi2 tests compared the distributions, determined by manual and automated means, of grades in the omitted screens and of grades in the screens following omitted screens. If p<0.05 we planned post-hoc analyses to compare the distributions of each grade versus all others with Holm-Bonferroni correction (8 tests).
RESULTS: There were 135194 subjects with 677871 screens in the test set (5 per subject). The automated ND assessment had 91.0% sensitivity and 82.6% specificity for detection of ND by manual assessment. Manual and automated ND assessment would omit 154818 and 159008 screens of which 1043 and 1006 had more-than-mild disease (MTMD) respectively. The distribution of disease by manual and automated ND assessment in the omitted (p<10−8) and following (p <10−3) screens are shown in the table. There was no difference for individual grades of MTMD.
CONCLUSION: A DR screening policy with 2YSI would have reduced screens by 23% and missed 2.9% screens with MTMD. Allocation of 2YSI by automated means is safe regarding incident MTMD in the 2YSI and is efficient in that it omitted a further 0.6% of screens. Therefore this automated DR grader could also perform screening interval allocation.
