**Background:** Glaucoma is the leading cause of irreversible blindness worldwide, with a projected prevalence of 111.8 million by 2040. Accurate and timely diagnosis and management are critical. Biofluid markers (e.g., from serum, tears, aqueous humour) have been explored to understand glaucoma pathogenesis, but the complexity of interactions between biomarkers requires advanced analytical strategies. Artificial intelligence (AI) and bioinformatics tools, including supervised techniques (e.g., artificial neural networks, discriminant analysis) and unsupervised methods (e.g., cluster analysis, principal component analysis), have shown promise in analyzing these complex data. This systematic review aimed to describe the application of AI and bioinformatics in the analysis of biofluid markers in glaucoma, appraise the evidence for clinical implementation, and identify areas for future research.
**Methods:** The review was conducted in accordance with PRISMA guidelines and registered in PROSPERO (CRD42020196749). A comprehensive search was performed across five electronic databases (Embase, Medline, Cochrane Central Register of Controlled Trials, Cochrane Database of Systematic Reviews, and Web of Science) from inception to August 11, 2020, updated August 1, 2021. Inclusion criteria were original peer-reviewed studies analyzing biomarker concentrations using AI/bioinformatics to predict or modify therapy/outcome/diagnosis in glaucoma or elevated intraocular pressure, with samples from vitreous fluid, aqueous fluid, tear fluid, plasma, serum, or ophthalmic biopsies. Exclusion criteria included pediatric diseases, non-human studies, post-mortem samples, non-English publications, reviews, and studies using only regression analysis without application to treatment or prognosis. Two independent reviewers screened abstracts and full texts, with disagreements resolved by a third reviewer. Data extraction was performed by one reviewer with 10% verified by a second. Risk of bias was assessed using Joanna Briggs Institute Critical Appraisal Tools, with studies classified as low (>80% yes), moderate (50-79% yes), or high risk (<49% yes). Narrative synthesis was undertaken due to heterogeneity.
**Key Results:** The search retrieved 10,258 studies, and 39 met inclusion criteria. Study designs included 23 cross-sectional (59%), nine prospective cohort (23%), six retrospective cohort (15%), and one case-control (3%). Primary open angle glaucoma (POAG) was the most common subtype (55%). Twenty-four studies examined disease characteristics, 10 explored treatment decisions, and five provided diagnostic clarification. Serum was the most common biofluid (54.5%), followed by aqueous humour (36.3%). Over 175 unique differentially expressed biomarkers were reported, with only nine biomarkers implicated in POAG by multiple studies (e.g., glutamine, referenced in three studies). Predictive accuracy of AI models ranged from 51% to 95%, with artificial neural networks generally most accurate. Sensitivity ranged from 81-90% and specificity from 87-93%. Area under receiver operating curve (AUROC) values ranged from 0.58 to 0.93, with most >0.85. For example, Barbosa Breda et al. (2020) reported AUROC of 0.91 (LDA) and 0.93 (SVM) for differentiating glaucoma from controls. Beutgen et al. (2019) achieved sensitivity 81% and specificity 93% (AUROC 0.875). Grus et al. (2008) reported sensitivity 90% and specificity 87%. However, biomarker selection significantly affected accuracy; Pan et al. (2020) had AUROC 0.62 using d-erythronalactone 2 but 0.86 with galactose 1. Quality appraisal showed 16 studies with low risk of bias, 18 moderate, and 5 high. Common limitations included small biofluid volumes (20-200 μL), lack of reporting on non-significant biomarkers, and black-box AI models without rationale for algorithm selection.
**Clinical Implications:** AI analysis of biofluid markers demonstrates strong diagnostic potential, with models achieving accuracy comparable to or exceeding human diagnosis using imaging (e.g., 90.0% and 94.8% accuracy reported by Yang et al. 2019). However, no clear pathogenic mechanism has emerged due to heterogeneity in biomarker findings. The tools have not been tested in clinical contexts, and implementation faces barriers including need for specialized expertise, financial costs, and integration into workflow. Future studies should validate AI models in diverse populations, compare AI diagnostic accuracy to gold-standard techniques, and ensure complete reporting of methods and populations. Despite limitations, AI-driven biofluid analysis could augment existing imaging-based tools for glaucoma diagnosis and management.