Patients with sight impairment (previously referred to as partial sightedness) or severe sight impairment (previously referred to as blindness) are eligible for a certificate of visual impairment (CVI). Certification confers eligibility for a wide range of benefits administered by local authorities but requires registration by a consultant ophthalmologist.
National certification criteria are guided by visual acuity and perimetry data but registration is ultimately a clinical decision, and guidance is somewhat ambiguous and subjective. As a result of this lack of clarity, a large number of eligible patients do not receive a CVI, thereby missing out on social support in the community which can be crucial for maintaining their quality of life [1-4].
As a group of diseases, glaucoma represents the most frequent cause of irreversible blindness globally, and the second leading cause of sight impairment in England and Wales [5,6]. However, it is possible that British data are affected by imprecision generated by poor and uneven rates of CVI provision. Reasons for missed registration are unclear and very likely multifactorial, but patients from lower sociodemographic backgrounds have been shown to be at higher risk of missed registration [1,3,4]. In addition, patients that qualify on the basis of perimetry data rather than visual acuity are more likely to be missed, perhaps due to the above issues of ambiguity and subjectivity [2].
We sought to investigate the current rate of missed registration of glaucoma patients with blindness, but this required automation of the time-consuming and laborious process of thoroughly assessing visual field results.
The glaucoma field defect classifier
There are well validated and accepted criteria for defining levels of severity of visual field defects, many of which are specific to glaucoma. Amongst the best known are the Hodapp-Parrish-Anderson criteria, which can be mapped to CVI registration criteria to solve the issues of ambiguity and subjectivity. However, the criteria are too laborious for routine use in clinic: it requires counting squares on pattern deviation plots and cross-checking against mean deviation and central field decibel levels [7].
We developed a Glaucoma Field Defect Classifier (GFDC) to automate implementation of Hodapp-Parrish-Anderson criteria. It uses a basic computer vision application to identify loci on pattern deviation plots and count these (Figure 1). Moreover, as it does not leverage machine learning, the algorithm is fully interpretable and not liable to changes or unpredictable performance drops. To initially validate the algorithm, we cross checked outputs against judgements made by clinicians manually assessing perimetry plots, finding that GFDC has perfect accuracy across levels of visual field defect [8].
Early clinical deployment
We deployed our semi-automated approach for determining CVI eligibility using GFDC on a cohort of patients attending tertiary glaucoma clinics over a 12-month period [9]. Some 37% of patients with blindness were not registered for a CVI, mapping closely to previous estimates over the last 30 years [1-4]. While GFDC identified seven patients that were not CVI-eligible on further investigation, this was because these patients did not have data from a better-seeing eye available. This reaffirmed the accuracy of GFDC and demonstrated its potential to support identification of CVI-eligible but missed patients.
We explored potential reasons for missed registration and concluded that multifactorial explanations were common (Figure 2). However, four common themes emerged through our investigation:
- Administrative failure or lack of consent: where patients were correctly identified and referred but were either not offered or declined CVI.
- Co-morbidity and frailty: where patients were cared for by multiple specialties, had many co-morbid health problems, or documented severe frailty.
- Reversibility: where patients were planning to undergo treatment that was anticipated to improve their vision e.g. cataract surgery.
- Mental health diagnoses including schizophrenia and delirium-on-dementia.
Figure 1: Under the bonnet of the GFDC: A simple computer vision algorithm identifies loci on the pattern deviation plot and incorporates counts with the mean deviation and the central global plot decibel values to classify glaucomatous field defects as ‘no defect’, ‘mild’, ‘moderate’, or ‘severe’. It does not involve machine learning, maximising interpretability and reliability while also minimising computational costs.
Figure 2: Results from a cross-sectional study of 5620 glaucoma patients in Cambridge, with semi-automated analysis powered by the GFDC. Sixty-four severely sight impaired patients were identified by semi-automated screening, of whom seven had data from a better-seeing eye missing. Of 57 truly severely sight impaired patients, 21 were not registered for a CVI. Potential reasons for missed certification were diverse and was frequently multifactorial. Adapted from Thirunavukarasu, et al. in British Journal of Ophthalmology [9].
Future uses
The most obvious use case for GFDC is to support initiatives to improve CVI provision for eligible glaucoma patients [2]. The application does not require expert use to operate, although a consultant may be required to sign off on certification due to current legal requirements. Future adaptations of our approach could incorporate criteria from different jurisdictions around the world (as definitions of blindness and sight impairment vary) or adjust the perimetry algorithm for use with diseases other than glaucoma.
Other use cases follow on from how perimetry classification criteria are frequently used in other contexts [10]. Clinical research is of significant interest, including recruitment for clinical trials where eligibility criteria are restricted to patients with certain levels of severity of disease. In the future – if studied appropriately – patients may have more individualised treatment and follow-up based on the severity of their disease, with apps like GFDC used for automatic classification. For instance, surgery may be undertaken sooner for patients with more severe disease, or follow-up intervals extended for patients with consistently mild disease.
The code can be adapted to work with perimetry plots at scale (‘batch processing’), although for many perimetry devices this can still require manual export of individual test results due to manufacturer-imposed limitations. We are working with specific devices to automate data input to minimise the friction of using GFDC. Conclusion The GFDC is freely hosted online for clinicians and researchers to use. Those interested in using the algorithm at scale can also access the source code freely. Arun J Thirunavukarasu, Rohan Sanghera, and Federico Lattuada are able to help with deployment and troubleshooting.
References
1. Barry RJ, Murray PI. Unregistered visual impairment: is registration a failing system? Br J Ophthalmol 2005;89(8):995–8.
2. King AJW, Reddy A, Thompson JR, Rosenthal AR. The rates of blindness and of partial sight registration in glaucoma patients. Eye 2000;14(Pt4):613–9.
3. Bunce C, Evans J, Fraser S, Wormald R. BD8 certification of visually impaired people. Br J Ophthalmol 1998;82(1):72–6.
4. Robinson R, Deutsch J, Jones HS, et al. Unrecognised and unregistered visual impairment. Br J Ophthalmol 1994;78(10):736–40.
5. Quartilho A, Simkiss P, Zekite A, et al. Leading causes of certifiable visual loss in England and Wales during the year ending 31 March 2013. Eye (Lond) 2016;30(4):602–7.
6. GBD 2019 Blindness and Vision Impairment Collaborators & Vision Loss Expert Group of the Global Burden of Disease Study. Causes of blindness and vision impairment in 2020 and trends over 30 years, and prevalence of avoidable blindness in relation to VISION 2020: the Right to Sight: an analysis for the Global Burden of Disease Study. Lancet Glob Health 2021;9(2):e144–60.
7. Hodapp E, Parrish RK, Anderson DR. Clinical Decisions in Glaucoma. Mosby; Maryland Heights, USA; 1993:52–61.
8. Thirunavukarasu AJ, Jain N, Sanghera R, et al. A validated web-application (GFDC) for automatic classification of glaucomatous visual field defects using Hodapp-Parrish-Anderson criteria. NPJ Digit Med 2024;7(1):131.
9. Thirunavukarasu AJ, Jain N, Yu Helmut CY, et al. Semi-automated screening reveals patients with glaucoma-induced blindness missing out on social support: a cross-sectional study of certificate of visual impairment allocation. Br J Ophthalmol 2025:bjo-2024-326745.
10. Ng M, Sample PA, Pascual JP, et al. Comparison of visual field severity classification systems for glaucoma. J Glaucoma 2012;21(8):551–61.
