Sampling for quality assurance of grading decisions in diabetic retinopathy screening Designing the system to detect errors

摘要

Purpose - To evaluate various designs for a quality assurance system to detect and control human errors in a national screening programme for diabetic retinopathy. Design/methodology/approach - A computer simulation was performed of some possible ways of sampling the referral decisions made during grading and of different criteria for initiating more intensive QA investigations. The effectiveness of QA systems was assessed by the ability to detect a grader making occasional errors in referral. Findings - Substantial QA sample sizes are needed to ensure against inappropriate failure to refer. Detection of a grader who failed to refer one in ten cases can be achieved with a probability of 0.58 using an annual sample size of 300 and 0.77 using a sample size of 500. Originality/value - An unmasked verification of a sample of non-referrals by a specialist is the most effective method of internal QA for the diabetic retinopathy screening programme. Preferential sampling of those with some degree of disease may improve the efficiency of the system.