If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Academic Unit of Ophthalmology, Institute of Inflammation & Ageing, College of Medical and Dental Sciences, University of Birmingham, UKOphthalmology Department, University Hospitals Birmingham NHS Foundation Trust, Birmingham, UK
INSIGHT Health Data Research Hub for Eye HealthNIHR Biomedical Research Centre At Moorfields Eye Hospital NHS Foundation Trust, UCL Institute of Ophthalmology, London, UK
Academic Unit of Ophthalmology, Institute of Inflammation & Ageing, College of Medical and Dental Sciences, University of Birmingham, UKOphthalmology Department, University Hospitals Birmingham NHS Foundation Trust, Birmingham, UKHealth Informatics, University Hospitals Birmingham NHS Foundation Trust, Birmingham, UK
Corresponding author: Susan P Mollan, Ophthalmology Department, University Hospitals Birmingham NHS Foundation Trust, Birmingham, UK, , +44 (0)1213716612
Ophthalmology Department, University Hospitals Birmingham NHS Foundation Trust, Birmingham, UKHealth Informatics, University Hospitals Birmingham NHS Foundation Trust, Birmingham, UK
This article describes the data contained within the INSIGHT Birmingham, Solihull and Black Country Diabetic Retinopathy Dataset.
Design
Dataset descriptor for routinely collected eye screening data.
Introduction
Diabetic Retinopathy (DR) is the most common microvascular complication associated with Diabetes Mellitus (DM), affecting approximately 40% of this patient population. Early detection of DR is vital to ensure monitoring of disease progression and prompt sight saving treatments as required. The INSIGHT Health Data Research Hub for Eye Health is a National Health Service (NHS)-led ophthalmic bioresource which provides researchers with safe access to anonymised routinely collected data from contributing NHS hospitals to advance research for patient benefit.
Methods
This report describes the INSIGHT Birmingham, Solihull and Black Country DR screening dataset, a dataset of anonymised images and linked screening data derived from the UK’s largest regional DR screening programme.
Participants
All diabetic patients aged 12 years and above, attending annual digital retinal photography-based screening within the Birmingham, Solihull and Black Country Eye Screening Programme
Main outcome measures
This dataset consists of routinely collected data from the eye screening programme. The data primarily includes retinal photographs with the associated diabetic retinopathy grading data. Additional data such as corresponding demographic details, information regarding patients’ diabetic status and visual acuity data is also available. Further details regarding available data points is available in the supplementary information, in addition to the INSIGHT webpage included below.
Results
At the time point of this analysis (31/12/2019) the dataset comprised 6,202,161 images from 246,180 patients, with a dataset inception date of 01/01/2007. The dataset includes 1,360,547 grading episodes between R0M0 and R3M1.
Conclusion
This dataset descriptor article summarises the content of the dataset, how it has been curated and what its potential uses are. Data is available through a structured application process for research studies that support discovery, real world evidence analyses and innovation in artificial intelligence technologies for patient benefit. Further information regarding the data repository and contact details can be found at https://www.insight.hdrhub.org/.
Population trends in the 10-year incidence and prevalence of diabetic retinopathy in the UK: a cohort study in the Clinical Practice Research Datalink 2004-2014.
The Birmingham, Solihull and Black Country Eye Screening Programme, which commenced in 2007, aims to invite all diabetic patients aged 12 years and above for annual digital retinal photography-based screening. The programme fulfils the national Public Health England Diabetic Eye Screening Programme requirements and encompasses patients who fall under nine hospital catchment areas across Birmingham, Dudley, Walsall and Wolverhampton, within the West Midlands (England, UK), and is hosted by University Hospitals Birmingham NHS Foundation Trust (UHB) to which the screening data is uploaded. This screening programme provides screening to over 200,000 individuals and includes longitudinal follow up data up to 15 years; it is thought to be one of the largest urban diabetic screening schemes in Europe.
The INSIGHT Hub aims to maximise the benefits and impact of historical, patient-level NHS hospital admission and electronic health record data by making it research-ready including curation, pseudonymisation and anonymisation. INSIGHT is one of a number of Health Data Research Hubs established by UK Research & Innovation (UKRI) through Health Data Research UK (HDRUK). INSIGHT was formed through a collaboration partnership between the NHS (University Hospitals Birmingham NHS Foundation Trust (UHB) and Moorfields Eye Hospital NHS Foundation Trusts (MEH)), academia (University of Birmingham), industry (Roche, Google), and charity (Action Against Age-related Macular Degeneration; AAAMD). INSIGHT enables access to anonymised routinely collected patient data from UHB and MEH, focusing on eye health, and the emerging field of ‘Oculomics’ in which the eye is used as a ‘window’ into systemic health, including discovery of novel biomarkers for diseases such as dementia and ischaemic heart disease.
Built on the ethically-approved INSIGHT Research Database, the hub has established efficient and robust governance processes that support safe and secure access to anonymised extracts of ‘evergreen’ datasets that are continuously updated in line with the clinical services. One of these datasets is the INSIGHT Birmingham, Solihull and Black Country DR dataset,
a research-ready longitudinal record of routinely collected screening data relevant to diabetic eye disease.
In this article we describe the INSIGHT Birmingham, Solihull and Black Country DR dataset by creating a datasheet which utilises the headings of ‘motivation, composition, collection process, preprocessing/cleaning/labelling, uses, distribution, and maintenance.’ This format is adapted from the datasheets for datasets guidance, outlined by Gebru et al, and included all sections relevant to the INSIGHT Birmingham, Solihull and Black Country DR dataset.
Gebru T, Morgenstern J, Vecchione B, Vaughan JW, Wallach H, Daumé H III, et al. Datasheets for Datasets [Internet]. arXiv [cs.DB]. 2018. Available from: http://arxiv.org/abs/1803.09010
The pathogenesis of DR involves microangiopathy and capillary occlusion leading to retinal ischaemia and an increase in vascular endothelial growth factor levels. As a result, macular oedema and retinal neovascularization are responsible for sight loss.
Although laser and surgical interventions such as panretinal and focal retinal photocoagulation are available for advanced neovascular DR, prevention of disease and its progression is vital.
In conjunction with preventative measures, early detection and prompt treatment of DR is important to minimising visual loss. The St. Vincent Declaration in 1989 stated that a primary objective for Europe should be a reduction in diabetes-related blindness by at least one-third.
In response to the high burden of diabetes-associated ocular morbidity, the NHS Diabetic Eye Screening programme for England was initiated in 2003 with the primary objective being to reduce sight loss among the diabetic population through early detection.
Advances in computing power and the field of machine learning have introduced new avenues of research in healthcare, particularly in diagnostic systems. The use of artificial intelligence (AI) for the detection of DR has been illustrated in the literature, with studies showing promising evidence for a potential transition automated screening in the future.
Feasibility and patient acceptability of a novel artificial intelligence-based screening model for diabetic retinopathy at endocrinology outpatient services: a pilot study.
A key barrier to AI training and validation are a shortage of datasets containing sufficient volumes of data with reference standards and accurate labelling.
In addition to AI development, datasets can be utilised for the development of other novel diagnostic and interventional technologies. The main advantages of DR data include:
●
Large volumes of routinely collected longitudinal data obtained from an ethnically diverse population representing an entire region of England.
●
Accurate ground truth data from a nationally endorsed screening programme with robust processes for participant inclusion and quality management of data.
2.2 Dataset composition
The Birmingham, Solihull and Black Country Diabetic Eye Screening Programme is set within the West Midlands. The region includes a diverse ethnic, and socio-economic mix, with a higher than UK average of minority ethnic groups. There are particularly high rates of diabetes, physical inactivity, obesity, and smoking in this region.
The INSIGHT Birmingham, Solihull and Black Country DR dataset is routinely collected data from the eye screening programme and comprises data relating to multiple episodes for each patient. Each episode (patient visit) includes the retinal photograph along with corresponding demographic and DR grading data. The dataset contains all instances collected via the eye screening programme from its inception to present day. Images are graded to output a retinopathy and maculopathy score as shown in table 2. A grade is assigned for retinopathy (R) and maculopathy (M), for example (R0M0) would signify that the patient has no DR at the time the images were graded. Patients receive a grade based on the eye with the most advanced DR and are directed into the appropriate pathway.
Table 2Table showing routine digital screening grades by year of screening including retinopathy (R) and maculopathy (M) grades. Incomplete grading data includes grades with a retinopathy label but no maculopathy label and vice versa. Inadequate images are those where the image itself may be of reduced quality due to things such as media opacities (for example cataracts), but the grading can still occur. Ungradable images are those where grading was considered inappropriate as the images were so poor (for example due to patient factors such as movement etc). In 2014 the new grades of R3S (Stable) and R3A (Active) were introduced. Values in the DR (Any Grade) column marked with an asterisk is not included in the totals column on the right hand side of the table. This column contains totals of all grades from R1M0 to R3SM1.
The dataset has been kept updated with new patient encounters via the screening programme. Table 1 illustrates demographic data representing the patients who have been seen in the screening programme (based on all cases registered between 01/01/2007 to 31/12/2019). The age of participants displayed in table 1 is the age at first encounter (first visit screening appointment with the service). Over the past few years a drive has been made to reduce missing ethnicity data by obtaining ethnicity from primary care systems for active patients. In currently available data, 87% have ethnicity recorded (with 13% being unknown).
Table 1Demographic data for patients included within this dataset. Demographic details included here are those recorded at patients’ first presentation. For example if a patient was aged 30 at their first attendance to the screening service, then they would be in the 21-30 age bracket in the table above.
The analysis described here reflects a time-locked data extract as at 31/12/2019 comprising data collected between [01/01/2007] and [31/12/2019], hereafter referred to as INSIGHT Birmingham, Black Country and Solihull DR Dataset Extract 2007-2019. From 2007 to 2019, the dataset includes 6,202,161 images from 246,180 individuals. Figures 4 and 5 show a breakdown of patients and images by year. Key data included in the dataset are:
●
Total number of patients screened and graded over a 13 year period.
●
Demographic information (including age, sex and patient reported ethnicity)
●
Diabetes status
●
Diabetes type
●
Length of time since diagnosis of diabetes
●
Visual acuity
●
The national screening diabetic screening grade category (seven categories from R0M0 to R3M1)
●
Diabetic eye clinical features
●
Reason for sight impairment and severe sight impairment
●
Screening Outcome (digital surveillance and time; referral to Hospital Eye Service (HES)).
Figure 2Bar graph showing the total number of images obtained through the eye screening programme per year from 2007 to 2019.
Figure 3Bar graph showing screening outcome distribution per year 2007-2019. This graph does not include patients whose outcome was annual recall (routine digital screening). Screening outcomes are described in further detail in section 2.3.2.
Figure 4Bar chart showing the availability of longitudinal data. The number of screening occurrences is plotted against the number of patients. For example, there are 12669 patients for whom data relating to 10 screening visits is available.
A full description of data points that are available are listed in supplementary file 1. An online data dictionary is available at: https://web.www.healthdatagateway.org/dataset/36886b21-12ff-45e7-82bc-fb5308c12450. There are a number of data points which are essential to be reported (such as demographics, visual acuity, grading, triage outcome). Some fields are mandatory to fill such as grade, and other fields have optional manual input such as diabetes type. The screening service data system has evolved over the years of operation, with the addition and removal of fields; these fields are therefore limited in their completeness. Datapoints are linked longitudinally between patient encounters to enable tracking of disease progression.
Grading labels are assigned through the eye screening programme and this process is described in more detail in the next section. The INSIGHT Birmingham, Black Country and Solihull DR Dataset Extract 2007-2019 comprises 1,360,547 grading episodes between R0M0 and R3M1, with 44,335 grading episodes having incomplete grading data, inadequate images, ungradable images or no photos. Table 2 shows the composition of the data set by DR grade and year.
The dataset is composed of routinely collected data which represents real-world clinical processes within the eye screening programme. Given that the screening service aims to offer appointments to all patients in the area with DR over the age of 12 years there is a high level of inclusion, although it is recognised at a national level that there is reduced engagement with the screening service by certain groups and that, despite being a free health service, some groups may experience barriers to access such as transport, caring or work duties or the need for a carer to accompany them.
Whilst this dataset is self-contained, INSIGHT provides the capability to provide linked health data on those individuals within this dataset who also attend UHB, whether for hospital eye care or for other systemic health issues. This more holistic record can support discovery of new associations between retinal status and systemic health in people with diabetes. Researchers can therefore apply to access datasets that represent patient care within the eye screening programme alone, or patient care combining the screening service and their linked UHB hospital eye care
INSIGHT Health Data Research Hub for Eye Health. UHB Linked Diabetic Eye Disease from National Screening to Hospital Eye Care [Internet]. Health Data Research Innovation Gateway. 2021 [cited 2021 Dec 11]. Available from: https://web.www.healthdatagateway.org/dataset/7c087b3c-cc04-4d3d-894b-7ea5419fa714
INSIGHT Health Data Research Hub for Eye Health. UHB Linked Diabetic Eye Disease from National Screening to Hospital Eye Care [Internet]. Health Data Research Innovation Gateway. 2021 [cited 2021 Dec 11]. Available from: https://web.www.healthdatagateway.org/dataset/7c087b3c-cc04-4d3d-894b-7ea5419fa714
INSIGHT Health Data Research Hub for Eye Health. UHB Linked Diabetic Eye Disease from National Screening to Hospital Eye Care [Internet]. Health Data Research Innovation Gateway. 2021 [cited 2021 Dec 11]. Available from: https://web.www.healthdatagateway.org/dataset/7c087b3c-cc04-4d3d-894b-7ea5419fa714
As an example when the diabetic retinopathy screening data was linked to the UHB hospital cardiac outcomes, the DR screening-cardiac outcome dataset comprised 1,760,093 eye images from 61252 individuals with 272863 ICD-10 cardiac codes and 10693 SNOMED cardiac episode codes.
Who. The international statistical classification of diseases and health related problems ICD-10: Tabular list v. 1. 2nd ed. Genève, Switzerland: World Health Organization; 2005. 1200 p.
The collection process is described as it applies to the Birmingham, Solihull and Black Country Eye Screening Programme 2007-2019 dataset, and as it continues to be undertaken at the time of writing (01 September 2022); it cannot be assumed to continue unchanged beyond this date since this will reflect both local and national guidance. The collection process is summarised in four sections. The first focuses on the historical development of the screening service and screening pathways; the second outlines the types of screening categories that differ from routine digital screening annually; the third describes image capture; the fourth describes the grading (image labelling) pathway.
2.3.1 Birmingham, Solihull and Black Country Diabetic Eye Screening Programme
The Diabetic Eye Screening Programme started in 2007, centred on Birmingham Heartlands Hospital, with a total of 6 hospitals participating in 2007, an additional hospital joining in 2008, and two Black Country hospitals joining in 2014 (one of which on-boarded retrospective data from 2007). The screening programme operates on an optometry based model, with nine contributing hospital eye services. Ninety sites were originally involved; this has increased with time to 110 screening sites.
All patients with diabetes aged above 12 years are offered yearly diabetic eye screening nationally. Over 130,000 patients were screened in 2019 with almost 80% of eligible patients responding to an invitation for screening. If the person has no retinopathy, they are returned to annual screening, however if they have pre-proliferative or active proliferative retinopathy, patients are referred on for further assessment and treatment in hospital eye services.
2.3.2 Screening Outcomes
There are five screening outcomes in this dataset aside from those in routine digital screening annually:
●
Surveillance pathway for DR
●
Surveillance with slit-lamp biomicroscopy
●
General Practice (GP) Non DR (Those referred back to primary care)
●
Ophthalmology DR
●
Ophthalmology Non DR
Surveillance screening is additional in the year screening for any patient who has changes in the retina, changes that may develop or disappear in due course, but if left for 12 months could potentially lead to a change in the management plan such as the need for treatment. A patient would have their normal annual screen and be referred through to Surveillance rather than hospital eye service for extra attention within the screening year. The surveillance category was developed in Birmingham by Professor Paul Dodson in the early 2000s and recommended by the National Screening Service in 2010. In effect it allows closer monitoring of patients in the hospital setting without the need to refer through to the ophthalmology clinics in the hospital eye service, thereby reducing the hospital eye service workload and being a more cost effective measure. Surveillance clinics are staffed by trained screener/graders and the images graded by senior graders, with options including return to annual screening, screen again within a determined number of months, and refer to hospital eye service.
Surveillance with slit-lamp biomicroscopy usually follows an annual screening instance where images are ungradable (U); ungradable images are ones where it is inappropriate to grade as the images are so poor (for example due to patient factors such as movement). A transition period between 2014 -2015 can be observed where there was an increase in the number of ungradable images (Table 2). This was due to a change in the process which meant that it was impossible to send a patient from annual screening without submitting images. If the image was ungradable often an anterior shot would be taken, providing evidence that the equipment was not dysfunctional. This process enabled reimbursement of the screening visit. Surveillance with slit lamp takes place within a hospital eye service clinic by senior screeners trained to perform dilated fundoscopy using slit lamp biomicroscopy; it is purely a clinical assessment by the screener, with no images being recorded.
GP Non DR These patients are those who are seen in the Diabetic Eye Screening Programme, but do not have DR and so are referred back to primary care for future care.
Ophthalmology DR, are those who are referred to hospital eye services for ophthalmology appointments. Those in the category for Ophthalmology Non DR, are those who are referred to the hospital eye services for other eye conditions such as suspected glaucoma, cataract etc.
Within the active patients in the screening programme in the year 2019, patient outcomes included 121,667 for annual recall. 9,033 patients were sent to the surveillance pathway for DR; 2,302 patients were sent to slit-lamp biomicroscopy surveillance; 2,920 patients were referred to Ophthalmology for DR; 1,209 were referred to Ophthalmology for non-DR ophthalmic conditions; and 788 patients were referred back to their GPs due to not having DR (Figure 1). A proportion of patients have the screening outcome ‘other.’ The ‘other’ category includes patients with screening outcomes that do not fit in the five outcome categories. Outcome data for these patients (other) is unavailable.
Figure 1Bar graph showing the number of patients seen per year from 2007 to 2019.
Pregnancy DR screening data is also available within this dataset and is available to be requested from INSIGHT. The national requirement is to offer screening after the first antenatal clinical appointment.
If diabetic retinopathy is detected, then an additional retinal assessment is offered between 16 to 20 weeks. Another retinal assessment is recommended at 28 weeks with referral onward to the hospital eye service as necessary.
2.3.3 Image capture
Images are captured using retinal photography and transferred to the centralised grading centre at Birmingham Heartlands Hospital, part of University Hospitals Birmingham NHS Trust. The majority of the retinal cameras in use are Topcon cameras (TRC-NW6), with other cameras including Canon CR-DGi, Nidek AFC-210 and Kowa Alpha 8. Once images have been captured, they are ingested into OptoMizeTM (Digital Healthcare Limited, United Kingdom) and stored in a SQL database hosted by a client server within UHB.
2.3.4 Grading pathway
The grading pathway is supported by OptoMizeTM, and is conducted using a ‘feature-based’ assessment. Prior to 2018, images were screened by graders who made decisions regarding DR status. Following this, graders are now responsible for identifying and entering retinal features into the OptoMize software, which uses a rule based decision system to generate the grading outcome. Retinal photographs are queued by the patient identifier in the software, and screened by graders chronologically. If patients are deemed to have background DR (R1) or worse, then the images are forwarded to a second grader who will identify features independently of the first grader. The OptoMize software then decides whether patients will need to be referred to a more experienced arbitration grader, who will see the first two grading outcomes and make an informed judgement on the final grade outcome. Additionally, quality control measures ensures 10% of patients with no DR will be referred to a secondary grader. This grading process, involving manual validation and quality control, ensures that data capture is accurate.
This grading pathway, as part of the screening programme, generates data labels for the image set within the INSIGHT Birmingham, Solihull and Black Country DR Dataset. The image set would include at least one disc-centred and one macula-centred image of adequate quality per eye. It is important to note that the grade is assigned per eye rather than by the individual image. This means that grades cannot currently be provided on a per image basis. The final referral outcome is based on the worst grade of the two eyes.
2.4 Ethics
The INSIGHT Birmingham, Solihull and Black Country DR Dataset is created through the INSIGHT Research Database, which was approved by the West of Scotland Research Ethics Committee 4 in 2020 (Ref: 20/WS/0087)), and received all institutional governance approvals in the same year. A commitment to use routinely collected data in anonymised form ‘to support research and improve care for others’ is enshrined within the NHS Constitution.
INSIGHT is one of a number of UK initiatives that supports this within a strict governance framework and with patient and public involvement to provide independent oversight of the data access processes.
but in brief: (1) INSIGHT respects the request of any patients who do not wish their data used in this way, and has robust processes working with NHS Digital to ensure that no data is included within the research database from individuals who have exercised their right to opt out using the NHS Digital National Data Opt-Out (NDOO) service.
The associated NHS trusts actively promote awareness of their proposed use of the data amongst their patient communities, and of the option to opt out (including instructions on how to do so). (2) Independent review of Data Use Applications to INSIGHT is conducted by the INSIGHT Data Trust Advisory Board (DataTAB) which comprises independent membership of patients, public and sector experts; the DataTAB advisory recommendation to the Data Controller (in this case UHB) is critical to any decision to provide data access. (3) Patients and the public are also involved in the processes informing the development of the INSIGHT Hub and through two Lay Advisors who are members of the INSIGHT Leadership Group.
2.5 Preprocessing/cleaning/labelling
Images are processed using INSIGHT cloud-based technology in pairs, consisting of a full-sized image along with the corresponding thumbnail. Full images contain EXIF data and are high resolution, whereas the thumbnails are smaller sized files with no EXIF data. Full images are pushed through EXIF processing and through conversion to DICOM (or other required image format) using Moorfields Librarian, a custom-built software tool created by Softwire Technology Limited (London, UK) for Moorfields Eye Hospital NHS Foundation Trust. This custom-built tool is not open source and the code is owned by softwire. Due to the volume of data being transferred into the UHB environment, quality control was integrated into the pre-processing pipeline. The filetype and content of the image was validated using automated tooling developed for pre-processing, so each image is individually validated. Thumbnails are pushed through two Deep Learning (DL) classifiers developed within UHB: (1) the ‘Sorting Hat Model’ and (2) the ‘Laterality and Fixation Model’.
2.5.1 Thumbnail processing
Identifying retinal images: The Sorting Hat Model is a deep learning classifier that was developed to distinguish anterior eye images and posterior retinal fundus images (AUC>99.9%), in order to provide a pure retinal image dataset for diabetic retinopathy studies The model is written in Python 3.8 and built on the TensorFlow 2.0 framework.
Identifying laterality and fixation: Laterality and fixation are useful to researchers but are not mandatory fields for the DR screening service, and are therefore not always recorded by the human graders. Laterality is recorded in approximately 65% of images and fixation in approximately 7% of images. INSIGHT constructed the Laterality and Fixation Model which runs on the retinal images output by the Sorting Hat Model. This model is used to label the image laterality (left eye vs right eye; AUC of 99.08%), and whether the retinal image was disc-fixated or macular-fixated (AUC 99.27%). This model is written in Python 3.8 and built on the TensorFlow 2.0 framework.
2.5.2 Full image processing
EXIF processes: Over 95% of full-sized images are obtained in Joint Photographic Experts Group (JPEG) format. They are first run through the EXIF processes. All EXIF processes use EXIFTOOL which is publicly available (https://exiftool.org/). The “EXIF grabber” reads and stores all of the EXIF data. The EXIF tags for images include: camera information, camera setting information and image settings. The “EXIF stripper” is then used to strip all EXIF data from the images, to remove any non-essential unique data and so reduce any risk of re-identification.
Image conversion (including DICOMisation): The next step in full image processing involves use of the Librarian application, which converts images from JPEG into DICOM format. Although most images ingested into the INSIGHT hub from the DESP are in JPEG format, some images are formatted to PNG. The PIL package in Python is used to convert PNGs to JPEGs using quality at 75 (PIL default), before the images are converted to DICOM using Librarian. As the images are validated on ingress, the main quality control on egress is to ensure that DICOM images remain uncorrupted. Manual validation of the librarian software was completed during the testing and development phase, in additional to the initial production phase. Quality assurance is also performed on images during extraction for fulfilment of a data request.
The INSIGHT Birmingham, Solihull and Black Country DR Dataset is available in both JPEG and DICOM formats. Image pixel data is not altered during the processes described above.
2.6 Uses
The INSIGHT Birmingham, Solihull and Black Country DR Dataset has been prepared to support research for patient benefit including from discovery to validation using both tabular and image data. Specific examples include: discovery of novel associations with diabetic retinopathy; exploration of health disparities; analysis of trends over time; development of artificial intelligence as a medical device (AIaMD) tools for DR classification; validation of these tools across populations, including as part of regulatory applications. In addition, when linked to systemic data, applications include: identification of retinal biomarkers of systemic disease in the context of diabetes including: cardiovascular or cerebrovascular event; renal failure; peripheral vascular disease; peripheral neuropathy; foot ulcers; anaemia; dementia; or other systemic health output that would be routinely collected within routinely collected hospital data. Publications using INSIGHT data will be cited on the website to illustrate examples of data use. The dataset description may be viewed (and access applied for) on the INSIGHT website and via the HDRUK Innovation Gateway.
Health Data Research UK. INSIGHT: Eye Health [Internet]. Health Data Research Innovation Gateway. 2020 [cited 2021 Dec 20]. Available from: https://web.www.healthdatagateway.org/collection/6524818791020588
Data Use Applications for the INSIGHT Birmingham, Solihull and Black Country DR Dataset can be made via the HDRUK Innovation Gateway or by contacting the Hub directly (www.insight.hdrhub.org). The Data Use Application form includes description of the researcher, a plain English summary of the project, the expected public benefit and detailed description of how the data will be used and the data use environment. Applications are welcomed from all bona fide researchers representing recognised research organisations with a clear commitment to patient benefit. All research data applicants should be able to demonstrate information security and health data research best practice.29,30,31,32,33,34,35
Data Use Applications undergo sequential stages of evaluation: first, internal INSIGHT checks, including due diligence (applicant/institution) and evaluation of whether the dataset is suitable for the proposed project; second, review by the DataTAB, providing independent advisory recommendation regarding the use of data and anticipated patient and public benefit; third, evaluation by the Data Controller (in this case UHB) who has the legal responsibility and makes the final decision.
If the application is supported, the Data Controller and the applicant proceed to contractual discussions, including agreeing access arrangements and financial terms that secure a sustainability and fair value return to the NHS. The contractual discussions normally take the form of establishing a Data Licence Agreement.
The Data Controller may determine how access to the data is made available. In this case, through the UHB Trusted Research Environment (TRE), or an equivalently secure data environment that has been approved by the Data Controller. The TRE is a provisioned cloud environment. Data cannot be downloaded for local use and APIs are not available for data access. The data is provided in DICOM format for retinal images, and CSV for tabular data. JSON and YAML formats are not available.
In addition to complying with all UK GDPR and Data Protection law and best practice, the INSIGHT Data Use Application process aligns to the “Five Safes” framework:
Safe data: data is treated to protect any confidentiality concerns.
2.
Safe projects: research projects are approved by data owners for the public good.
3.
Safe people: researchers are trained and authorised to use data safely.
4.
Safe settings: a secure research environment prevents unauthorised use.
5.
Safe outputs: screened and approved outputs that are non-disclosive.
2.8 Strengths and limitations
Strengths of this dataset (and indeed other datasets available through INSIGHT) are their scale, their richness and their diversity. There are a number of publicly accessible DR datasets available globally.
Furthermore, the majority of these datasets are not routinely updated with real world data. The INSIGHT dataset contains over 6 million images with up to 15 years of follow up data. Lastly, linked systemic data can be requested through INSIGHT, allowing for investigation of new associations between eye and systemic health.
The main limitations of the INSIGHT dataset are those common to real world datasets, including the completeness of data and the level of quality assurance when compared to datasets from well-conducted clinical trials; however, this is mitigated by the dataset being derived from a quality-assured diabetic screening service with a high level of verification of image labels. An additional limitation is that there is no pixel specific annotation of images. In comparison, Porwal et al describe pixel level annotation data, allowing those without specialist retinal expertise to train algorithms.
The dataset extraction described here runs up to 31st Dec 2019. The COVID-19 pandemic had a significant impact on health services globally, including the UK.
Mollan SP, Fu DJ, Chuo CY, Gannon JG, Lee WH, Hopkins JJ, et al. Predicting the immediate impact of national lockdown on neovascular age-related macular degeneration and associated visual morbidity: an INSIGHT Health Data Research Hub for Eye Health report. Br J Ophthalmol [Internet]. 2021 Sep 13; Available from: http://dx.doi.org/10.1136/bjophthalmol-2021-319383
The DR screening service was modified in response to the pandemic, and patients that were in annual recall in 2019 (graded R0M0) were not screened in 2020, with a screen scheduled for 2021. There are plans ahead for a potential 2 year recall pathway for routine screening. In light of the impact of the pandemic and subsequent changes to the pathway, we limited the data extract reported here to the period 2007-2019 inclusive. Data after this period is also available for research purposes through INSIGHT but is not described in this paper.
3. Summary
This article describes the INSIGHT Birmingham, Solihull and Black Country DR dataset, including a detailed description of the 2007-2019 data extract
. This dataset is a large-scale, updating anonymised data resource generated securely from routinely collected NHS data (specifically the Birmingham, Solihull and Black Country Diabetic Eye Screening Programme). The dataset comprises over 6 million retinal photographs and relevant longitudinal clinical data, with capability to include other relevant systemic health data where appropriate. Access to the dataset is provided by the Data Controller (UHB) through the INSIGHT Health Data Research Hub. This datasheet provides a summary of the dataset to encourage transparency in dataset creation and the development of novel technologies, in addition to enhancing communication between dataset creators and users. Further information and contact details can be found at https://www.insight.hdrhub.org/.
Funding INSIGHT is a National Health Service-led partnership established to improve healthcare by encouraging research using routinely collected eye data. INSIGHT is funded by Health Data Research, United Kingdom (HDR UK). HDR UK is funded by the Medical Research Council (MRC), Engineering and Physical Sciences Research Council (EPSRC), Economic and Social Research Council (ESRC), National Institute for Health Research (NIHR); Chief Scientist Office of the Scottish Government Health and Social Care Directorates, Health and Social Care Research and Development Division (Welsh Government; HCRW), Public Health Agency (Northern Ireland; PHA), British Heart Foundation (BHF) and the Wellcome Trust Limited as trustee of the Wellcome Trust.
10 Data Security Standards [Internet]. Digital Social care. 2019 [cited 2021 Dec 20]. Available from: https://www.digitalsocialcare.co.uk/data-security-protecting-my-information/national-policy/
Population trends in the 10-year incidence and prevalence of diabetic retinopathy in the UK: a cohort study in the Clinical Practice Research Datalink 2004-2014.
Gebru T, Morgenstern J, Vecchione B, Vaughan JW, Wallach H, Daumé H III, et al. Datasheets for Datasets [Internet]. arXiv [cs.DB]. 2018. Available from: http://arxiv.org/abs/1803.09010
Feasibility and patient acceptability of a novel artificial intelligence-based screening model for diabetic retinopathy at endocrinology outpatient services: a pilot study.
INSIGHT Health Data Research Hub for Eye Health. UHB Linked Diabetic Eye Disease from National Screening to Hospital Eye Care [Internet]. Health Data Research Innovation Gateway. 2021 [cited 2021 Dec 11]. Available from: https://web.www.healthdatagateway.org/dataset/7c087b3c-cc04-4d3d-894b-7ea5419fa714
Who. The international statistical classification of diseases and health related problems ICD-10: Tabular list v. 1. 2nd ed. Genève, Switzerland: World Health Organization; 2005. 1200 p.
Health Data Research UK. INSIGHT: Eye Health [Internet]. Health Data Research Innovation Gateway. 2020 [cited 2021 Dec 20]. Available from: https://web.www.healthdatagateway.org/collection/6524818791020588
10 Data Security Standards [Internet]. Digital Social care. 2019 [cited 2021 Dec 20]. Available from: https://www.digitalsocialcare.co.uk/data-security-protecting-my-information/national-policy/
Mollan SP, Fu DJ, Chuo CY, Gannon JG, Lee WH, Hopkins JJ, et al. Predicting the immediate impact of national lockdown on neovascular age-related macular degeneration and associated visual morbidity: an INSIGHT Health Data Research Hub for Eye Health report. Br J Ophthalmol [Internet]. 2021 Sep 13; Available from: http://dx.doi.org/10.1136/bjophthalmol-2021-319383
This dataset descriptor illustrates the INSIGHT DR screening dataset, which is a dataset of anonymised images and linked screening data from the UK’s largest regional DR Screening Programme.
Declaration of interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Professor Keane has received speaker fees from Heidelberg Engineering, Topcon, Carl Zeiss Meditec, Haag-Streit, Allergan, Novartis and Bayer. He has served on advisory boards for Novartis and Bayer and has been an external consultant for DeepMind and Optos.
Professor Mollan reports consulting fees from Invex Therapeutics and Neurodiem; advisory board compensation from Janssen, Santhera, GenSight Biologics and Chugai-Roche Ltd; and speaker fees from Chiesi, Heidelberg engineering, Allergan, Chugai-Roche Ltd, Santen, Santhera, Roche and Teva. All conflicts of interest are outside the submitted work.
00025-8&id=gr2.jpg){kind=link}
00025-8&id=gr3.jpg){kind=link}
00025-8&id=gr4.jpg){kind=link}
00025-8&id=gr1.jpg){kind=link}