figure a

Introduction

Traditionally, artificial intelligence (AI) refers to the ability of computing systems to recognize patterns and mimic human cognitive features (e.g., generalize and learn from experience) in high volumes of data [7]. Machine learning (ML) is a type of AI that informs extraction of generalized principles from data, by using algorithms comprised of explicit instructions about the data represented as mathematical models [8]. Although the line between ML models and traditional statistical models (e.g., logistic regression) is not well-defined, distinctions between simple AI or sophisticated ML models (i.e., complex AI) have been proposed. ML differs from traditional statistical approaches, hereafter referred to as “simple AI,” in that ML is programmed to learn from examples rather than being programmed with rules [9]. Bioinformatics involves both storing and analysis of biological information using computer technology [10]. In proteomics, which involves the analysis of the expression levels of large numbers of proteins and identification of their biological function and pathways [10], ML techniques can be used to (1) identify the functional significance of resultant proteins using Gene Ontology [11] (a process referred to as enrichment analysis); (2) scrutinize putative biological pathways using comprehensive pathway databases including Kyoto Encyclopedia of Genes and Genomes (KEGG) [12]; and (3) interactively map protein–protein interaction networks by referring to large data mining repositories such as STRING [13]. In metabolomics, computational methods are used to obtain information about endogenous and exogenous metabolites in various tissues as to assess pathophysiological changes and disease-related metabolic pathways using sources such as MetaboAnalyst software [14, 15]. These techniques have provided the opportunity for exploration into the molecular events involved in the development of RVO and generate large datasets of data that can be further analyzed using AI tools, which are well-suited for the extraction of useful information from these biological data [16]. As such, an increasing number of ophthalmology studies have performed bioinformatics in conjunction with AI tools using biofluids as biomarkers.

Exploration of biofluids using AI and bioinformatics may provide insight for the development of more targeted future therapies for retinal diseases such as RVO. Therefore, we set out to systematically review the literature describing applications of AI and bioinformatics-based analyses using biofluids as biomarkers in RVO. Additionally, this comprehensive review provides a summary and appraisal of both the methodology and conclusions of eligible studies with a focus on evaluating the potential of clinical implementation of these technologies.

Methods

This systematic review was conducted in accordance with the Preferred Reporting Items for a Systemic Review and Meta-analysis (PRISMA) guidelines [17]. The complete protocol for the study is available on PROSPERO (reg. CRD42020196749). This systematic review is focused on retinal occlusive diseases and is part of a series of systematic reviews on AI/bioinformatic applications in ophthalmology using biofluids.

Search strategy

A comprehensive search of five databases, including Embase, MEDLINE, Cochrane Central Register of Controlled Trials, Cochrane Database of Systematic Reviews, and Web of Science from inception through August 11, 2020, was conducted. An update of the search strategy was undertaken on August 1, 2021. The search strategy was developed to include the following medical subject headings (MeSH) derived from three categories including ophthalmology terms, AI/bioinformatics terms, as well as proteomics, metabolomics, and lipidomics (Appendix 1). No language or study design restrictions were applied to the search strategy.

Inclusion and exclusion criteria

All studies pertaining to intra-ocular ophthalmic conditions (i.e., affecting the anatomy or function of the internal structures or surface of the eye) investigating the role of biofluid marker concentrations to predict outcomes, disease conversion/prognosis, disease etiology, and risk factors or to modify patient treatment plans using AI/bioinformatic analyses were included. Biofluid samples extracted from vitreous fluid, aqueous fluid, tear fluid, plasma serum, or ophthalmic biopsies from patients enrolled in the study were considered eligible. Studies that combined biofluid markers with other markers such as imaging, demographics, and genetics in their statistical analyses were acceptable provided that they included at least one biofluid that was a type of protein, cytokine, lipid, or metabolite.

Studies were excluded based on the following criteria: (1) articles that referred exclusively to eye diseases that only affect pediatric patients (e.g., retinopathy of prematurity), (2) studies on non-human subjects (animal or in vitro cell studies), (3) studies including post-mortem biofluid samples, (4) cross-sectional, simple regression analyses without an application of their findings in the study’s populations, and (5) abstracts, reviews, systematic reviews, meta-analyses, single-case reports, editorials (without adequate study details and data presentation), and any type of non-peer-reviewed article.

Finally, the subset of studies that met the inclusion criteria and referred to retinal occlusive diseases was selected for the current review.

Study selection and quality assessment

All studies were screened by two independent reviewers (DRP, AP) first by title and abstract, followed by full-text screening. All disagreements were adjudicated by a third independent reviewer (SK and TF). Data extraction was performed by one reviewer using a standardized data abstraction form. To ensure completeness and consistency of our methodology, throughout the data extraction process, 10% of extractions were randomly checked by a second independent reviewer.

Risk of bias (ROB) and quality assessment of the identified studies were performed using the Joanna Briggs Institute Critical Appraisal Tools (JBI) [18, 19]. For each article, JBI criteria questions were noted as “yes,” “no,” “unclear,” or “not applicable.” The assessment was performed by one reviewer (DRP), and none of the studies were excluded from the review. Studies that reached up to 49% of questions as “yes” were classified as high ROB; from 50 to 69% as moderate ROB; and more than 70% as low ROB [20].

Data synthesis

Given the heterogeneity in biofluids, AI tools used, and study designs, a meta-analysis was not performed. Means and standard deviations (SD) were used to describe the age of the study populations. For each article, study design, location, type of RVO, biofluids examined, sample size, sex (proportion of males to females), and study aim were tabulated. Additionally, studies were categorized based on the type of statistical model, AI, or bioinformatic analyses used. The purpose of the methodology was also noted.

Results

Study characteristics

The search strategy identified 10,264 articles after removal of duplicates (Fig. 1). Of the 14 articles included in the current study, 7 were prospective (50%), 3 retrospective (21%), and 4 cross-sectional (24%) studies (Table 1). The country of origin of the studies spanned globally, with the majority from Japan (5, 54%) and China (4, 29%). The studies included 578 individuals with RVO (342 BRVO, 191 CRVO, 9 hemi-CRVO, 36 RVO type unspecified) and 201 controls. The mean age was 62.6 (SD = 5.1), and 46% of participants were female. Most studies aimed to identify an optimal treatment strategy (9; 54%) or the pathophysiology underlying the disease (4; 29%).

Fig. 1
figure 1

PRISMA flowchart diagram for study identification and selection. The PRISMA flow diagram for the systematic review presenting the number of studies included and excluded at each screening step, and reasons for exclusion

Full size image
Table 1 Summary characteristics of the included studies
Full size table

Appendix Fig. 2 provides details on risk of bias assessment. Overall, three articles were deemed to have a high [6, 21, 22], four moderate [3, 5, 23, 24], and seven low risk of bias [4, 18, 25,26,27,28,29]. Three of the four cross-sectional studies assessed did not identify confounding factors [3, 4, 6], and two did not use objective, standard criteria for measurement of RVO [5, 6]. Most of the cohort studies (n = 5, 66%) did not identify confounding factors nor statistical strategies for addressing them [21,22,23].

Characteristics of biofluid markers

The majority of biofluids were obtained from the aqueous humor (9, 64%), followed by vitreous humor (3, 21%), and only two used serum or whole blood (Table 2). The number of biofluids (514 in total across all studies) reported varied between 94 in RVO including CRVO and BRVO (Clusterin, Complement C3, Ig lambda-like polypeptide 5, Opticin, Vitronectin, etc.) [3] and 49 proteins, mainly implicated in angiogenesis, oxidative stress, and collage synthesis (fibroblast growth factor-4, alpha crystallin A chain, etc.) [6], in BRVO.

Table 2 Characteristics of biofluids, and artificial intelligence/bioinformatics analysis categorization and purpose
Full size table

Biofluid markers involved in pathogenesis

A metabolomic analysis [4] of contributing proteins to CRVO pathophysiology identified 37 out of 248 metabolites related to aberrations in amino acid metabolism, carbohydrate, and fatty acid metabolism in the aqueous humor of CRVO patients compared to controls (i.e., cataract patients). Total plasma homocysteine (HCY) during fasting and low vitamin B12 levels were associated with an increased risk of RVO, especially CRVO [25], while post-methionine load test HCY, serum folate, and methylenetetrahydrofolate reductase (MTHFR) mutation were not. Zeng et al. [5] reported 29 out of 39 vitreous chemokines (IFN-γ, IL-1β, IL-2, IL-4) to have a statistically higher concentration in CRVO or BRVO complicated with unresolved or condensed vitreous hemorrhage compared to patients with idiopathic preretinal membranes (PRMs) and idiopathic macular holes (IMHs) and 3 chemokines, IL-8, CXCL9, and TNF-α concentrations being more than six times higher in the RVO group compared to control (i.e., PRMs and IHMs patients). Another proteomic study found 5 out of 94 measured vitreous proteins to differ between RVO (CRVO, hemi-CRVO, BRVO) and control group [3] (idiopathic floaters). Clusterin, Complement C3, IGLL5, Opticin, and Vitronectin were suggested to be involved in the inflammatory response pathway, complement activation, and coagulation cascade.

Biofluid markers involved in secondary complications

A proteomic study on 6 patients [6] suggested the implication of 49 proteins related to angiogenesis, oxidative stress, and collagen synthesis in BRVO with macular edema. Other studies found reported various inflammatory factors such as vitreous IL-6 and VEGF [23], aqueous humor baseline platelet-derived growth factor-AA (PDGF-AA) [27], IL-8, and VEGF-A [24], to be related to pathogenesis of BRVO with macular edema. Patients with vitreous hemorrhage secondary to ischemic RVO were found to have 29 increased concentrations of chemokines in their vitreous, with IL-8, CXCL9, and TNF-a showing the greatest increase compared to controls (PRMs and IMHs).

Biofluid markers indicated in treatment response

In patients with BRVO, macular edema improvement in VA and decrease in foveal thickness after pars plana vitrectomy were predicted by vitreous IL-6 levels [23]. Four other studies showed that the level of certain aqueous humor biofluids were associated with treatment response in BRVO macular edema. Age, baseline central macular thickness (CMT), and baseline aqueous PDGF-AA predicted the number of intravitreal injections needed to control macular edema [27], and IL-8, chemo-attractant protein 1, and soluble VEGF receptors predicted visual gain and CMT reduction at 6 months of intravitreal anti-VEGF injections [24]. Higher pre-treatment aqueous IL-12 was found to be associated with inferior response to bevacizumab treatment (defined as CMT recovery < 90% at 1 month after treatment) [21]. Aqueous flare, another factor predictive of intravitreal anti-VEGF injections treatment of BRVO macular edema, was correlated with aqueous levels of soluble VEGFR-1, PlGF, MCP-1, sICAM-1, IL-6, and IL-8. Similarly, 11 out of 27 measured cytokines (VEGF, receptor antagonist interleukin-1, IL-6, IL-8, IL-9, IL-10, IL-12r70, IL-13, IL-15, etc.) were significantly different compared to cataract patients and significantly changed after intravitreal anti-VEGF injections both in CRVO and BRVO [22]. Finally, a prospective observational study on 31 patients reported that a smaller improvement in visual acuity (LogMAR units) after vitrectomy for macular edema in patients with CRVO was associated with higher VEGF and lower pigment epithelium-derived factor (PEDF) vitreous levels compared to those with a better outcome [29].

Applications of AI and bioinformatics

Studies varied in their reporting of AI algorithm development, ranging from simply describing the standard analysis protocol of established bioinformatics software, such as Gene Ontology analysis (6), STRING database (5), MASCOT search engine (6), Proteome Discoverer with SEQUEST algorithm (3), to dividing the protein data into a discovery and testing stage and performing additional validation analyses using receiver operating characteristic curves (ROC) [3]. Four articles performed proteomic analyses [3, 5, 6, 22], two of which integrated complex AI tools such as discriminant analysis [22] (prediction and classification), probabilistic clustering [3], and string pathway analysis [5] for identification of pathophysiology or biomarker discovery. One study performed metabolomics in conjunction with AI tools for clustering (principal component analysis), classification, and predictive modeling (orthogonal partial least squares discriminant analysis) of contributing proteins to CRVO pathophysiology [4]. There were nine studies that employed simple AI, logistic, and linear regression, to predict various outcomes in RVO using biofluid biomarkers. Compared to the complex AI papers, which were mainly aimed at investigating pathogenesis, most of the simple AI papers investigated cytokines, chemokines, or other biofluids sampled from aqueous, vitreous, serum, or blood, as predictors of treatment response.

Discussion

This is the first systematic review, to our knowledge, to describe the applications of AI and bioinformatics-based analyses using biofluids as markers in RVO. We evaluated the potential of these tools to aid clinical decision-making, specifically whether the levels of biofluid markers could predict visual prognosis [18, 23], treatment response [21, 28], and number of intravitreal injections needed to control RVO complications [22, 26, 27].

For example, the difference in aqueous humor cytokine concentration before and after intravitreal ranibizumab therapy for RVO may be used to predict which patients will have a more favorable response to the therapy and which will have an insufficient response [22]. Such cytokine profiling can be performed using proteomics and complex AI such as multifactorial discriminant analysis using Mahalanobis distance [22]. Bioinformatics approaches can make use the proteome of the aqueous humor to predict which patients may proceed to develop macular edema with BRVO [6].

Additionally, three of the four articles utilizing proteomic analyses proceeded to bring proteomic data into a functional context using bioinformatic tools such as SEQUEST as part of Proteome Discoverer (3), MASCOT search engine and GO analysis (6), and STRING database (5). Given the large output of data produced by proteomic analyses, bioinformatic tools are crucial in providing functional interpretation of protein–protein interactions in the case of STRING database, protein identification in the case of SEQUEST and MASCOT, and prediction of biological function using GO analysis. The results provided are probabilistic, and consequently the interpretation is constantly evolving as new experimental data is added to the bioinformatic databases. The risk of bias ranged from high in 21% to low in 50% of studies. The main methodological quality concerns were the lack of identification and strategies for confounding factors and not recruiting consecutive patients. We identified that operation of AI/bioinformatics as black-box models, lack of validation and testing, small sample volumes, and lack of healthy controls may serve as some of the challenges for their use in analysis of biofluids.

The reviewed studies used established AI tools such as discriminant analysis [22], probabilistic clustering [19], and principal component analysis [4], in conjunction with proteomic and metabolomic tools that are freely available and well-documented. However, only one study implemented a biomarker validation stage using ROC [3]. Cross-validation involves partitioning data into training, testing, and validation and provides estimations of the predictive value of the AI algorithm, which can reduce the bias associated with small sample sizes and the bias related to choosing the type of AI tools to use [30]. OCTs provide non-invasive near-microscopic visualization of retinal structures, are obtained relatively fast [31, 32], and can be used to support biofluid predictions. The combination of biofluids and OCTs has the potential to increase the clinical decision-making value of both techniques. For example, in BRVO, foveal thickness change post-vitrectomy has been correlated with vitreous IL-6 levels and found to predict best-corrected VA at 6 months [23], and central macular thickness and PDGF-AA were predictors of the number of intravitreal injections required to control macular edema [27]. In CRVO, anatomical response to anti-VEGF therapy, measured as central retinal thickness (CRT), was best predicted by baseline CRT and ICAM-1 [26]. Thus, individual-level measurements of cytokine expression correlated with OCT measures may help guide personalized treatment regiments that account for the fact that the rate of progression of disease is variable among patients.

Biofluid sampling is challenging due to small sample volume, inconsistent feasibility of office-based sampling, and lack of healthy controls. It was demonstrated that multiplex cytometric bead assay can reliably use a volume as small as 25 μl for cytokine analysis [33]. Half of the studies used enzyme-linked immunosorbent assay for sample analysis [6, 19, 21, 23, 24, 26, 29], a standard assay kit. One study [3] included in our analysis used a volume of 10 µl, and another did not mention the sample volume [21]. However, bias related to sample handling and processing must also be considered and could be reduced by appropriate protocol documentation and replication studies. While aqueous humor is more accessible in an office setting, vitreous samples are mostly obtained during vitrectomies. An alternative to vitrectomy samples is the collection of vitreous reflux after intravitreal injection using Schirmer’s tear strips, which can be done in the office [34]. Another challenge is that the interpretation of proteomic and metabolic studies may be confounded by the lack of healthy controls. Samples obtained from patients undergoing cataract surgery or intravitreal injections for other macular diseases may serve as good comparators with regards to biomarker levels [26, 35]. It is important to note, however, that the inflammatory markers identified in cases may be underestimated/overestimated when using comparators with and existing inflammatory state secondary to mild pathologies and/or exposure to procedures. Additionally, all of the samples from the assessed articles were taken after RVO had occurred, and therefore, it is not possible to definitely conclude if the biofluids identified are contributory to RVO and/or a reflection of the downstream consequences of RVO.

Changes in both serum and intraocular biofluid levels have been observed in RVO. None of the studies in the current review sampled the same biofluids in both serum and intraocular fluids. However, a consecutive case series, in which patients with proliferative vitreoretinopathy (PVR) and rhegmatogenous retinal detachment (RRD) were compared with macular hole (MH) or epiretinal membrane (ERM) patients, indicated that while the serum inflammatory profiles did not differ between groups, the concentrations of several cytokines were upregulated in PVR patients [36]. Additionally, concentrations of lipocalin-2 (LCN2) were found to be higher in the aqueous humor of CRVO patients compared to cataract patients, while no differences were noted in serum LCN2 levels [37]. These findings suggest that changes in serum biofluids, which largely indicate circulating systemic factors, have distinctive predictive significance compared to changes in the concentration of biofluids pertaining to the intraocular milieu. Nevertheless, as demonstrated in the current review, both serum and intraocular fluids may play a role in the identification of treatment strategy. Further studies are required to understand how to make use of the predictive value of the distinct profiles of serum and intraocular fluid inflammation-related factors of RVO patients in a clinical setting.

Conclusion

In this systematic review of 578 individuals and 514 biofluids, we documented the applications of AI and bioinformatics-based analyses using biofluids as markers in RVO etiology, prognosis, treatment response, and management. Overall, several studies have combined proteomics and metabolomics with AI analyses for clinical decision-making. Considering the limitations of these studies (e.g., lack of healthy controls, small sample sizes, small volumes), further validation of the studies outlined and comparison and integration of data obtained from various RVO severity, types, and treatment regiments are required before these techniques can be adopted for individual-level clinical treatment. In addition to these limitations, this review also highlights that although the application of AI and bioinformatics in RVO is poised to grow in the future, currently its use is only at its infancy.