top of page

Data Use Application 003

Lead applicant organisation name

Name of the legal entity that signs the contract to access the data.

King's College London (KCL).

Applicant: Paul Nderitu, Clinical Research Fellow, Ophthalmology Registrar and PhD Candidate.

Project title

The title of the project/research study request that the applicant is investigating through the use of health data.

Deep learning for the automated prediction of diabetic retinopathy progression.

Lay summary

A concise and clear description of the project.  This should outline the problem, objective and the expected outcomes in language that is understandable to the general public.

Diabetic retinopathy (DR) is a disease where the light-sensing layer at the back of the eye (retina) becomes damaged by raised blood sugar levels. DR is a complication of diabetes which affects 4.9 million people in the UK and 463 million worldwide. DR affects around one in three people with diabetes and is a leading cause of acquired vision loss in working-age adults. We developed deep learning (DL) models, a type of artificial intelligence, to predict when DR would reach a sight-threatening stage up to 3-years in the future using retinal images and participant characteristics from the southeast London DR screening service. The aim of the current study is to evaluate how developed DL models perform in a separate group of people with diabetes who have attended the diabetic eye screening programme in Birmingham, Solihull and the Black Country. Once evaluated, developed models could enable accurate, individualised risk predictions for people with diabetes who regularly attend DR screening services. This could mean fewer unnecessary visits for individuals at low-risk of DR progression with associated time and cost savings for the screening service, but also allow for closer monitoring and potentially earlier treatment for individuals at high-risk of DR progression, which could subsequently reduce their risk of avoidable vision loss.

Public benefit statement

A description in plain English of the anticipated outcomes, or impact of project on the general public.

Early detection of referable DR allows for the prompt referral to the hospital eye service for closer monitoring and treatment. DR screening is credited, in part, for the significant reduction in DR-associated severe sight impairment. However, it is estimated that some 80% of people with diabetes are at a low-risk of progression to sight threatening disease over 2 years. Our primary objective is to develop DL models that can predict individual-level risk of progression to referable DR or maculopathy over 3-years. The DL models would enable individualised risk-based follow-ups for all patients attending routine DR screening in the UK meaning low-risk participants could avoid unnecessary screening visits but those at risk can be detected early and referred for closer monitoring and treatment. If realised, individualised risk-based follow-ups could reduce screening appointments by 40% and screening costs by 20% with an estimated £23 saving per appointment from the individual/societal perspective. External validation of developed DL models would demonstrate their effectiveness in a separate DR screening service and hence the generalisability of developed models. This would pave the way for prospective safety and efficacy evaluation, which, if positive, could result in the use of developed DL models within the national DR screening service.

Latest approval date

The last date the data access request for this project was approved by a data custodian.

22 October 2022.

Dataset(s) name

The name of the dataset(s) being accessed.

Diabetic Eye Screening: the Birmingham, Solihull and Black Country.

Access type

Determines whether the data will be accessed with an Trusted Research Environment (TRE) or via data release.

Data released under Data Licensing Agreement to a secure Trusted Research Environment provided by the Data Controller University Hospitals Birmingham.

Data sensitivity level

The level of identifiability of the data being accessed, as defined by Understanding Patient Data.

Anonymised.

bottom of page