GlaukomAI: Clinical Validation of an AI System for Early Glaucoma Screening

Official title

GlaukomAI: Clinical Validation of an Artificial Intelligence-Based System for Early Glaucoma Screening and Diagnosis - A Case-Control Study and Referral Accuracy Assessment

Overview

Glaucoma is one of the leading causes of irreversible blindness worldwide. Early diagnosis is crucial to prevent vision loss, but current diagnostic pathways require multiple specialist visits and tests, leading to long waiting times and delayed diagnosis. This study aims to evaluate the accuracy of GlaukomAI, an artificial intelligence (AI)-based software that analyzes fundus photographs of the eye to detect glaucoma at an early stage. The study is conducted at IRCCS Fondazione G. B. Bietti (Rome, Italy) and is structured in two phases: * Phase 1 enrolls 200 participants (100 with diagnosed glaucoma and 100 healthy controls) to assess how accurately GlaukomAI can distinguish between glaucoma and healthy eyes, compared to the judgment of a panel of three expert glaucoma specialists. * Phase 2 enrolls 1,000 consecutive outpatients to evaluate whether GlaukomAI can correctly identify patients who need referral to a glaucoma specialist, and to compare its performance with that of non-specialist ophthalmologists. Participants undergo a single study visit including standard ophthalmic examinations (visual acuity, eye pressure measurement, visual field test, OCT, and fundus photography). No investigational drugs or invasive procedures are involved. The results of this study will provide evidence to support the integration of AI-based tools into routine glaucoma screening pathways, with the goal of reducing diagnostic delays and improving access to care.

Detailed description

Glaucoma is a chronic optic neuropathy representing one of the leading causes of irreversible blindness worldwide, with an estimated 111.8 million cases projected by 2040. Despite the availability of effective treatments, approximately 50% of affected individuals remain undiagnosed, as the disease progresses insidiously and symptoms often appear only when damage is already advanced and irreversible.

Current diagnostic limitations include high inter-operator variability in optic disc assessment, limited sensitivity of visual field testing in early stages, and suboptimal specificity of OCT (estimated at 72% in a Cochrane systematic review). No single examination provides sufficient diagnostic accuracy, accessibility, and cost-effectiveness for large-scale screening.

GlaukomAI (Sens-vue GlaukomAI) is an AI-based diagnostic software using deep learning with Convolutional Neural Network and Transformer architecture. It analyzes standard fundus photographs to detect key glaucoma biomarkers (neuroretinal rim appearance, inferior and superior sectors) and provides a diagnostic classification (Referable Glaucoma / Non-Referable Glaucoma) within 2-8 seconds per image. The system was trained on over 100,000 fundus images from diverse ethnicities, annotated by 30 eye care professionals and validated by 243 ophthalmologists and 208 optometrists across Europe.

Study Design

This is a prospective interventional clinical investigation with a non-CE-marked medical device, structured in two complementary phases:

* Phase 1 - Case-Control Diagnostic Accuracy Study: 200 participants (100 with diagnosed glaucoma, 100 healthy controls) are enrolled to assess the sensitivity and specificity of GlaukomAI against a gold standard defined by the consensus of a panel of three expert glaucoma specialists, based on multimodal assessment (fundus photography, OCT, and visual field). * Phase 2 - Prospective Referral Accuracy Assessment: 1,000 consecutive outpatients attending IRCCS Fondazione Bietti for any clinical reason are enrolled to evaluate the referral accuracy of GlaukomAI (binary output: Referable / Non-Referable) in a real-world setting, and to compare its performance with that of non-glaucoma-specialist ophthalmologists evaluating the same pseudonymized fundus images.

All participants undergo a single study visit (or two visits within one week if needed) including: best-corrected visual acuity measurement, slit-lamp biomicroscopy, Goldmann applanation tonometry, Humphrey visual field testing (24-2 SITA Standard or SITA Faster), fundus examination with Cup-to-Disc Ratio assessment, fundus photography using a widefield TrueColor Confocal imaging system (iCare DRS Plus), and retinal nerve fiber layer (RNFL) and ganglion cell layer (GCL+IPL) thickness assessment via Cirrus HD-OCT (Carl Zeiss). No investigational drugs or invasive procedures beyond standard clinical practice are involved.

Statistical Analysis For Phase 1, sample size was calculated to detect an expected sensitivity and specificity of 88% with 95% confidence and ±8% precision, yielding 100 subjects per group. For Phase 2, enrollment of 1,000 patients allows estimation of real-world sensitivity and specificity with ±5% precision, assuming a 10% glaucoma prevalence in a tertiary referral center. Both eyes will be included in the analysis using generalized estimating equations (GEE) or mixed-effects models to account for intra-subject correlation. Diagnostic performance metrics (sensitivity, specificity, PPV, NPV, AUC) will be calculated with 95% confidence intervals. Agreement between methods will be assessed using Cohen's kappa; comparisons will use McNemar's test.

Funding This study is funded under the Transforming Health and Care Systems (THCS) partnership, co-funded by the EU Horizon Europe Research and Innovation Programme (Grant Agreement No. 101095654).

Interventions

• Device GlaukomAI (Sens-vue GlaukomAI)
  GlaukomAI is an AI-based diagnostic software (Sens-vue ApS) that analyzes standard fundus photographs to detect glaucomatous changes. The system uses deep learning with Convolutional Neural Network and Transformer architecture to evaluate key glaucoma biomarkers, including neuroretinal rim appearance in the inferior and superior sectors. It accepts standard fundus images acquired with conventional fundus cameras or portable devices and provides a diagnostic classification (Referable Glaucoma / N

Primary outcome measures

• Diagnostic Accuracy of GlaukomAI - Sensitivity and Specificity [Time frame: At enrollment visit (single visit, or two consecutive visits within 1 week)]

Secondary outcome measures (2)

• Referral Accuracy of GlaukomAI vs. Non-Specialist Ophthalmologists [Time frame: At enrollment visit]
• Diagnostic Agreement - Cohen's Kappa [Time frame: At enrollment visit]

Eligibility criteria

Inclusion Criteria for all patients:

• Age >18 years
• Freely given informed consent obtained prior to study initiation
• The participant has the capacity to understand and the willingness to follow study instructions and is likely to complete all required visits and procedures

Inclusion Criteria for glaucoma patients:

Patients affected by any type of glaucoma (primary open-angle, primary angle-closure, secondary glaucoma) on pharmacological therapy

Inclusion Criteria for healthy controls:

• Absence of ocular pathologies
• IOP <21 mmHg
• Visual field and OCT within normal limits
• Optic disc of normal appearance on clinical evaluation

Exclusion criteria

• Presence of media opacities preventing the acquisition of adequate quality fundus imaging (e.g., advanced cataract, vitreous hemorrhage, severe corneal opacities)
• Retinal or optic nerve pathologies that could confound the diagnosis (e.g., non-glaucomatous optic neuropathies (ischemic, inflammatory, compressive), moderate-to-severe diabetic retinopathy, advanced macular degeneration, retinal vascular occlusions)
• Having undergone any ocular surgery in the past 3 months
• Inability to cooperate with perimetric examination

Criteria are shown verbatim from the registry (in English). Final eligibility is always assessed by the study center.

Healthy volunteers: Yes

Study design

Allocation

Non-randomized

Model

Single group

Masking

Double blind

Primary purpose

Diagnostic

Study locations

Center list to be confirmed — check the primary protocol.

Publications

• Li Z, He Y, Keel S, Meng W, Chang RT, He M. Efficacy of a Deep Learning System for Detecting Glaucomatous Optic Neuropathy Based on Color Fundus Photographs. Ophthalmology. 2018 Aug;125(8):1199-1206. doi: 10.1016/j.ophtha.2018.01.023. Epub 2018 Mar 2. PMID 29506863
• Lemij HG, Vente C, Sanchez CI, Vermeer KA. Characteristics of a Large, Labeled Data Set for the Training of Artificial Intelligence for Glaucoma Screening with Fundus Photographs. Ophthalmol Sci. 2023 Mar 17;3(3):100300. doi: 10.1016/j.xops.2023.100300. eCollection 2023 Sep. PMID 37113471
• Michelessi M, Quaranta L, Riva I, Martini E, Figus M, Frezzotti P, Agnifili L, Manni G, Miglior S, Posarelli C, Fazio S, Oddone F. Exploring the gap between diagnostic research outputs and clinical use of OCT for diagnosing glaucoma. Br J Ophthalmol. 2020 Aug;104(8):1114-1119. doi: 10.1136/bjophthalmol-2019-314607. Epub 2019 Nov 15. PMID 31732524
• Oddone F, Lucenteforte E, Michelessi M, Rizzo S, Donati S, Parravano M, Virgili G. Macular versus Retinal Nerve Fiber Layer Parameters for Diagnosing Manifest Glaucoma: A Systematic Review of Diagnostic Accuracy Studies. Ophthalmology. 2016 May;123(5):939-49. doi: 10.1016/j.ophtha.2015.12.041. Epub 2016 Feb 15. PMID 26891880
• Tham YC, Li X, Wong TY, Quigley HA, Aung T, Cheng CY. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology. 2014 Nov;121(11):2081-90. doi: 10.1016/j.ophtha.2014.05.013. Epub 2014 Jun 26. PMID 24974815
• Quigley HA, Broman AT. The number of people with glaucoma worldwide in 2010 and 2020. Br J Ophthalmol. 2006 Mar;90(3):262-7. doi: 10.1136/bjo.2005.081224. PMID 16488940

Identifiers

NCT: NCT07668193 · GLC 02-26 · 101095654

Primary sources (government registries)

• ClinicalTrials.gov (US NIH)

View this study on ClinicalTrials.gov ↗