Diabetic retinopathy screening in Australia: attitudes in primary care settings and the validation of artificial intelligence approaches

Aims The early recognition of diabetic retinopathy is critical in preventing visual morbidity and mortality. The aims were: • To examine the attitudes of General Practitioners (GPs) towards diabetic retinopathy (DR) screening in primary care. • To evaluate an algorithm for detecting DR developed by the Commonwealth Scientific and Industrial Research Organisation (CSIRO). • To evaluate the algorithm against two different ophthalmologist graders. • To evaluate the performance effect of re-training the algorithm. Methods • A qualitative analysis of interviews with 15 GPs in Australia. • Non mydriatic fundus photographs of participants in the LANDMark study, with and without diabetes, were graded and compared with the automated gradings of the CSIRO’s algorithm, after which the algorithm was re-trained and performance re- analysed. Results • 15 GPs aged 54.7±15.5 years completed interviews. None of the participants reported performing DR screening in their general practices. Seven themes were identified as key areas for the successful introduction of DR screening into wider practice. • The sensitivity of the CSIRO’s algorithm in detecting referable DR was 53.1%, 74.1% and 73.1% as compared with the first grader, the second grader, and where both graders agreed, respectively. The sensitivity of the re-trained algorithm was 59.4%, 85.2% and 84.6% as compared with the first grader, the second grader and where both graders agreed, respectively. Conclusion • This study identified specific strategies to enable the wider implementation of DR screening in general practice and to inform policymakers in developing a better framework for DR screening in primary care. • The measured performance of the algorithm depended on the reference standard used, and re-training the algorithm improved its performance. Grader variability, intrinsic errors within an algorithm’s ‘black box’ and good study design are all factors that can affect measured performance of an algorithm