Unveiling AI's role in papilledema diagnosis from fundus images: A systematic review with diagnostic test accuracy meta-analysis and comparison of human expert performance

Abstract

Background: Papilledema is a condition, which is characterized by optic disc swelling due to increased intracranial pressure. Diagnostic modalities include fundus camera and other ophthalmology imaging techniques. The Frisén scale is used to grade the severity of this condition. In this paper, we investigate the application of artificial intelligence (AI) for detecting and grading papilledema from fundus images.

Method: Following the PRISMA guidelines for systematic reviews, a search of five databases (PubMed, Scopus, Web of Science, Embase, Cochrane) was conducted using MeSH terms related to AI and papilledema. The inclusion criteria were original articles that discussed AI applications for detecting or grading papilledema from fundus images. Extracted data included sensitivity, specificity, accuracy, and technical and demographic characteristics.

Results: The systematic review included 21 studies. In the meta-analysis, the pooled sensitivity and specificity were 0.97 and 0.98, respectively. High heterogeneity was observed (I² > 96%). Deep learning models outperformed traditional machine learning algorithms, with detection models being more effective than grading models. Publication bias was observed with Deek's plot. Several publications compared AI to human experts, showing superiority or non-inferiority of computer algorithms to humans.

Conclusions: AI models show high diagnostic accuracy in detecting papilledema, often surpassing human experts in sensitivity, though not always in specificity. Despite limitations related to patient selection, image sourcing, and heterogeneity, AI holds potential to significantly improve diagnostic accuracy and clinical workflows in ophthalmology.

Keywords: Artificial intelligence; Deep learning; Fundus imaging; Machine learning; Papilledema.