Machine Learning Models Analyze Tabular Clinical Data to Improve Glaucoma Detection

A systematic review analyzing 35 studies published between 2010 and 2024 has evaluated how machine learning models perform when diagnosing glaucoma from structured tabular records, revealing diagnostic accuracies between 76 and 98.3 percent and area under the receiver operating characteristic curve values spanning 52.5 to 99 percent. Writing in BMC Biomedical Engineering, a research team led by Mohammad Hasan Shahriari, Farkhondeh Asadi, Hamid Moghaddasi, Arash Roshanpour, Farideh Sharifipour, and Zahra Khorrami examined algorithms trained on structured clinical measurements, ocular examinations, and patient demographics. The authors determined that integrating structured parameters, most notably optical coherence tomography measurements and visual field test metrics alongside demographic profiles, substantially enhances diagnostic capability compared to isolated evaluations. While deep learning architectures and support vector machines exhibited exceptional diagnostic potential across diverse datasets, the synthesized evidence also highlighted pressing methodological bottlenecks, including persistent dataset imbalances, sample size constraints, and a widespread reliance on retrospective private repositories that complicate broader clinical generalizability.

The clinical motivation driving the synthesis centers on the global burden and diagnostic complexity of glaucoma, a progressive cluster of ocular disorders defined by the degeneration of retinal ganglion cells and retinal nerve fiber layers leading to irreversible optic nerve damage and vision loss. Cited epidemiological figures indicate that glaucoma represents the second leading cause of blindness globally, impacting approximately 80 million individuals in 2020 and projected to reach 112 million by 2040. In the United States, data from the American Glaucoma Society shows that 2.7 million residents have glaucoma, yet only half are aware of their condition, contributing to annual healthcare and treatment costs near 2.5 billion dollars. Because early diagnosis is crucial to curbing permanent vision loss, clinicians rely on multifactorial examinations. However, conventional diagnostic workflows remain subjective, resource-dependent, and prone to variability, especially since individuals with normal-tension glaucoma experience damage despite maintaining intraocular pressure within statistically normal clinical limits.

To address diagnostic variability, artificial intelligence and machine learning frameworks have emerged as clinical decision support systems capable of detecting complex trends across vast volumes of patient records. While earlier diagnostic software in ophthalmology focused heavily on raw image processing across fundus photography and optical coherence tomography scans for diseases like diabetic retinopathy, diabetic macular edema, and keratoconus, interest has expanded toward structured tabular data in electronic health records. The systematic review highlighted previous foundational research demonstrating this potential, such as Raju and colleagues utilizing electronic health records from 650 healthcare facilities across the United States to develop a predictive analytics framework yielding an area under the curve of 0.80. Similarly, studies by Lee and colleagues combined demographic factors with optical coherence tomography data to achieve an area under the curve of 0.93, while Oh and coworkers established an interpretable model reaching an area under the curve of 0.94 that incorporated explanation mechanisms to clarify individual predictions.

To capture the landscape of tabular machine learning investigations, the researchers executed a literature search across five major databases: PubMed, Scopus, Web of Science, ScienceDirect, and the Institute of Electrical and Electronics Engineers database. Conducted on December 28, 2024, the search covered articles published between 2010 and 2024 using terms including Glaucoma, Artificial intelligence, Machine learning, Deep learning, Decision support, Decision aid, and Computer-assisted selected from medical subject headings. Strict eligibility criteria governed inclusion, requiring peer-reviewed English-language original research papers centered on glaucoma diagnosis or classification utilizing machine learning applied to structured tabular measurements such as demographics, clinical findings, optical coherence tomography variables, and visual field data. Studies relying solely on image data without tabular structures were excluded, alongside conference proceedings, book chapters, abstracts, editorials, gray literature, systematic reviews, meta-analyses, and investigations restricted purely to disease progression, treatment outcomes, or unrelated comorbidities.

The formal study selection followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 protocol, beginning with 10,590 retrieved records across the five databases. After 4,981 duplicate citations were removed using Endnote software, two authors, Mohammad Hasan Shahriari and Farkhondeh Asadi, independently screened the titles and abstracts of the remaining 5,609 articles, excluding 5,382 records. The same two reviewers assessed the full text of 227 articles, selecting 28 studies and incorporating seven additional articles identified from reference lists, resulting in 35 articles selected for analytical review. Data extraction was performed independently by the two primary reviewers, with any disagreements resolved through consultation with a third author, Hamid Moghaddasi. Methodological quality across the 35 included studies was formally evaluated using the Quality Assessment of Diagnostic Accuracy Studies 2 checklist across seven ordinal items evaluating risk of bias and applicability concerns.

Methodological evaluation via the Quality Assessment of Diagnostic Accuracy Studies checklist showed no severe bias regarding flow and timing, index tests, or reference standards, with only one study exhibiting a high risk of bias related to patient selection and no severe concerns regarding applicability. Dataset sizes across the 35 studies ranged from 100 records in a study by Eswari and colleagues to 33,611 records. The compiled databases reflected diverse clinical, paraclinical, and demographic data. Clinical examinations included visual acuity, refraction, slit lamp, tonometry, gonioscopy, and ophthalmoscopy. Structural examinations encompassed optical coherence tomography, optical coherence tomography angiography, pachymetry, corneal biomechanics, and fundoscopy, with retinal nerve fiber layer thickness and optic disc parameters frequently measured. Functional examinations involved visual field testing, notably mean deviation and pattern standard deviation values, while patient history captured age, gender, race, body mass index, and systemic diseases such as hypertension, diabetes, and atherosclerosis.

Data preprocessing emerged as an essential step across the reviewed investigations to enhance data quality, consistency, and model performance. Preprocessing workflows employed segmentation techniques to extract optic disc and cup regions from fundus images, along with normalization routines to place continuous features on comparable scales. Handling missing data involved excluding features with substantial missing values or applying imputation techniques. To address data imbalance between glaucomatous eyes and healthy controls, studies utilized techniques such as the Synthetic Minority Over-sampling Technique to balance datasets by generating synthetic samples, or class weight adjustments within loss functions. Feature selection proved equally vital for reducing dimensionality and eliminating noise. Investigators applied statistical tests such as t-tests and analysis of variance to isolate significant variables, as well as algorithmic approaches including recursive feature elimination and feature importance rankings derived from tree-based models, with comprehensive feature sets generally yielding superior performance.

Across the catalog of computational architectures, the systematic review observed extensive use of supervised machine learning algorithms, prominently led by support vector machines, deep learning models including artificial neural networks and multilayer perceptrons, random forests, and ensemble methods such as XGBoost, AdaBoost, Gradient Boosting, and Soft Voting. Additional supervised models included decision trees, logistic regression, naïve Bayes, and k-nearest neighbors, while less frequently used techniques encompassed partial least squares discriminant analysis, linear discriminant analysis, stochastic gradient descent, radial basis function networks, and relevance vector machines. Only one study utilized unsupervised machine learning, deploying k-means clustering to establish a glaucoma grading system based on visual fields. Hyperparameter tuning across these studies commonly relied on grid search and random search, whereas performance validation relied on k-fold cross-validation and leave-one-out cross-validation strategies to ensure robust internal evaluation across datasets. In most analyzed investigations, private institutional databases were utilized for model training and testing.

In evaluating overall diagnostic capability, reported performance metrics reflected notable efficacy across several algorithmic paradigms, with study accuracies spanning 76 to 98.3 percent, sensitivities ranging from 52 to 100 percent, specificities between 56 and 98 percent, and precision figures reaching 97 to 98 percent. Deep learning architectures trained on optical coherence tomography tabular data consistently demonstrated the highest discrimination, frequently achieving area under the curve scores surpassing 0.90, reflecting the rich structural feature sets extracted from retinal layers. In contrast, models relying exclusively on functional visual field parameters showed greater variability, with reported accuracies between 0.76 and 0.94, indicating that visual field measurements can vary across sessions due to patient cooperation, learning effects, and fatigue. Blending structural, functional, and demographic variables produced balanced diagnostic outputs, with studies combining visual field data, optical coherence tomography, and Synthetic Minority Over-sampling achieving sensitivity and specificity marks of 0.97 and 0.96.

Despite these favorable metrics, the review underscored critical limitations that currently restrict the clinical translation of machine learning algorithms in frontline care. Most evaluated models were trained on retrospective single-center databases, with external validation on an independent secondary database performed in only one identified study by Li and colleagues. Small sample sizes in multiple cohorts heightened the risk of overfitting, preventing models from learning generalizable patterns. Furthermore, the complexity of deep learning architectures poses interpretability hurdles for clinicians requiring transparent reasoning, highlighting the need for explainable artificial intelligence. The authors emphasized that future research demands prospective multicenter studies, standardized protocols for data acquisition and preprocessing, and the integration of systemic health records to track comorbidities like hypertension and diabetes. Addressing ethical priorities, including patient data privacy, encryption, equitable performance across diverse populations, and clear liability frameworks, will be essential for translating structured machine learning tools into clinical decision support systems.

Subject of Research: Systematic review of machine learning applications for glaucoma diagnosis using structured tabular data

Article Title: Applications of machine learning in glaucoma diagnosis based on tabular data: a systematic review

Article References: Shahriari, M. H., Asadi, F., Moghaddasi, H., Roshanpour, A., Sharifipour, F., & Khorrami, Z. (2025). Applications of machine learning in glaucoma diagnosis based on tabular data: a systematic review. BMC Biomedical Engineering, 7(1), Article 9. https://doi.org/10.1186/s42490-025-00095-3

Image Credits: AI Generated

DOI: 10.1186/s42490-025-00095-3

Keywords: Glaucoma Diagnosis, Machine Learning, Systematic Review, Tabular Data, Optical Coherence Tomography, Visual Field Testing, Deep Learning, Clinical Decision Support, Applications, machine, learning, glaucoma