Current applications of machine learning in the screening and diagnosis of glaucoma: a systematic review and meta-analysis

Authors: Murtagh P., Greene G., O’Brien C.

Geographical coverage: India, South Korea, Germany, United States of America

Sector: Glaucoma

Sub‑sector: Screening; diagnosis

Equity focus: Not reported

Study population: Patients with glaucoma

Review type: Effectiveness review

Quantitative synthesis method: Meta‑analysis

Qualitative synthesis method: Not applicable

Background

Glaucoma is a leading cause of permanent blindness worldwide, yet early detection is difficult because the disease is often asymptomatic. Global prevalence is projected to rise from 76 million cases in 2020 to 112 million by 2040. A common risk factor is elevated intra‑ocular pressure (IOP), which is usually treated with pressure‑lowering medication or surgery. Machine learning (ML) has become an important tool for glaucoma diagnosis, using algorithms to analyse imaging data. However, there are still challenges in validating these algorithms and comparing their accuracy across different diagnostic methods.

Objectives

To compare the diagnostic accuracy of two well‑established ML modalities—optical coherence tomography (OCT) and fundus photography—in the screening and diagnosis of glaucoma.

Main findings

Fundus photography and OCT demonstrated broadly comparable diagnostic accuracy. Fundus photography is, however, more accessible and cost‑effective for large‑scale screening.

The search identified 285 articles, of which 23 met the inclusion criteria and were incorporated into the quantitative synthesis (10 OCT studies and 13 fundus photography studies). Ten studies originated from Asia (nine from India and one from South Korea), and three used data from Germany or the United States.

In the random‑effects meta‑analysis, fundus photography had a pooled area under the receiver operating characteristic curve (AUROC) of 0.957 (95 % confidence interval [CI]: 0.917‑0.997; P < 0.001), while OCT had a pooled AUROC of 0.923 (95 % CI: 0.889‑0.957; P < 0.001). The marginally higher accuracy of fundus photography is likely attributable to the larger training dataset (59,788 images versus 1,743 for OCT).

Methodology
A literature search of PubMed and Embase was conducted on 1 February 2019 to identify observational studies that evaluated ML applied to fundus photographs or OCT images for glaucoma screening or diagnosis. Studies were excluded if they involved participants aged under 18 years, relied on human interpretation of fundus images, or based the ML algorithm on perimetry.

Two reviewers independently screened titles and abstracts. Reference lists of included studies were hand‑searched for additional publications. Data were extracted using a standardised form, and study quality was assessed with the Newcastle–Ottawa Scale. Findings were synthesised using inverse‑variance‑weighted meta‑analysis. A hierarchical summary ROC (HSROC) curve provided a single summary AUROC. Heterogeneity was evaluated with the I² statistic, and publication bias was assessed with funnel plots.

Applicability / external validity

Most included studies were conducted in Asia, with limited representation from other regions. The authors highlight the need for higher‑quality studies, noting variation in diagnostic criteria and lower methodological quality in several fundus photography studies.

Geographic focus

The included studies were conducted in India, South Korea, United States of America and Germany.

Summary of quality assessment

The review’s conclusions are of low confidence. Although searches were carried out in two major databases and bias was assessed using the Newcastle–Ottawa Scale, no grey literature searches, author contact, or expert consultation were reported. It is unclear whether two reviewers screened full‑text articles or extracted data, and language restrictions were not stated. Excluded studies were not listed, and results were not stratified by risk of bias.