MACHINE LEARNING APPLICATIONS IN OPTIC NERVE HEAD AND LAMINA CRIBROSA BIOMECHANICS: A SYSTEMATIC REVIEW AND META-ANALYSIS

Abstract

Introduction: In ophthalmology, artificial intelligence (AI) has become a revolutionary tool, especially for glaucoma detection and progression monitoring. However, the available data is still dispersed among many modalities and study designs. The lamina cribrosa (LC) and optic nerve head (ONH), which serve as the biomechanical entry point for retinal ganglion cell axons and support optic nerve stability, are crucial to this study. Glaucomatous optic neuropathy is closely associated with structural alterations in these areas, such as thinning or distortion. Even though cutting-edge imaging methods like MRI and OCT show promise, it is still challenging to adequately capture the intricate nonlinear biomechanical processes within microstructures using conventional techniques.

Methods: A thorough search covering research published between 2021 and 2025 was carried out in PubMed, Scopus, IEEE Xplore, and Web of Science. Peer-reviewed research using machine learning or deep learning for glaucoma diagnosis, progression prediction, or optic nerve biomechanics was included. Non-English articles, conference abstracts without full text, and research without quantitative results were excluded. Study design, modality, sample size, performance measures (AUC, accuracy, MAE, R2), and quality evaluation utilizing traffic-light bias matrices were the main topics of data extraction. Subgroup comparisons and pooled meta-analyses of diagnostic AUCs were included in the analytical synthesis.

Results: Twelve studies were found to be eligible. Biomechanical models (strain characteristics, ONH robustness) produced pooled AUC ≈0.84, whereas structural imaging models (OCT/OCTA, multimodal ensembles) produced pooled AUC ≈0.96. Thiéry 2023, Pourjavan 2024, and Lee EJ 2022 had good methodological quality, but smaller or retrospective studies had moderate to high bias, according to risk of bias evaluation. Although there was high heterogeneity (I² >90%), Begg's and Egger's tests revealed no significant publication bias.

Conclusion: AI shows exceptional glaucoma diagnostic accuracy, especially when combined with structural imaging. However, generalizability is limited by methodological inconsistency and heterogeneity. Multi-center validation established reference standards, and the incorporation of biomechanical insights should be given top priority in future research.