Diagnostic Accuracy of Current Machine-Learning Classifiers for Age-Related Macular Degeneration: A Systematic Review and Meta-Analysis

Author: Cheung R, Chun J, Sheidow T, Motolko M, Malvankar-Mehta M S

Geographical coverage: North America, Europe and Asia

Sector: Biomedical / Service delivery

Sub-sector: Diagnostic accuracy

Equity focus: Not reported

Study population: Patients with age-related macular degeneration (AMD)

Review type: Diagnostic-accuracy review

Quantitative synthesis method: Meta-analysis

Qualitative synthesis method: Not applicable

Background:

Age-related macular degeneration is a neuro-degenerative retinal condition and a leading cause of irreversible blindness worldwide, currently affecting about 170 million people; this figure is projected to rise to 288 million by 2040 owing to population ageing. Clinically, AMD is characterised by drusen and choroidal neovascularisation and is classified as either dry or wet. Although optical coherence tomography (OCT) and fundus photography provide reliable diagnosis, they are expensive and time-consuming, placing pressure on ophthalmology services. Artificial-intelligence methods—particularly machine learning (ML) classifiers such as convolutional neural networks and support-vector machines—offer the prospect of rapid, accurate and cost-effective automated diagnosis.

Objective:

To systematically review and meta-analyse the diagnostic accuracy of ML classifiers for all forms of AMD across available image datasets and to evaluate their suitability for clinical implementation.

Main findings:

Fourteen studies met the inclusion criteria; all contributed to the qualitative synthesis, and 13 (64 798 subjects; 612 429 images) were included in the meta-analysis. All were observational studies conducted in North America, Europe or Asia. Most datasets contained both wet and dry AMD, and four studies reported the disease stage explicitly.

Pooled sensitivity was 0.918 (95 % CI 0.678 – 0.980) and pooled specificity 0.888 (95 % CI 0.578 – 0.980). The relative odds of a positive test in AMD versus non-AMD eyes were 89.74 (95 % CI 3.05 – 2641.59). Positive and negative likelihood ratios were 8.22 and 0.09, respectively, indicating strong rule-in and rule-out potential. QUADAS-2 assessment showed low risk of bias for patient selection (64 %), index tests (71 %), reference standard (64 %) and flow/timing (100 %).

Methodology:

Searches of MEDLINE, EMBASE, CINAHL and ProQuest Dissertations & Theses, together with grey-literature sources, were carried out to 12 September 2020. Eligible studies employed advanced AI to diagnose any AMD type or stage from human ocular images (fundus photographs or OCT), reported diagnostic-performance metrics against a clinician-validated reference standard, and were published in English. Two reviewers independently screened records, extracted data and assessed quality using QUADAS-2; discrepancies were resolved by discussion. Random- or fixed-effects meta-analysis was applied according to heterogeneity (I², Chi-square, Z-value). Publication bias was explored with funnel plots.

Applicability / external validity:

Although the review highlights the promise of ML classifiers for tele-ophthalmology and service efficiency, generalisability is limited by heterogeneous imaging protocols, population differences and the retrospective nature of most datasets. No formal subgroup analyses were performed to assess these factors, and real-world performance in diverse clinical settings remains to be demonstrated.

Geographic focus:

Studies originated from North America, Europe and Asia; no geographical restrictions were applied.

Summary of quality assessment:

The comprehensive search strategy, duplicate screening and use of QUADAS-2 confer medium confidence in the findings. Nonetheless, English-language restriction, absence of an excluded-studies list, and limited exploration of heterogeneity temper the overall certainty.