Takahiro Ninomiya

Artificial intelligence holds the potential to transform ophthalmic assessment, yet its translation into clinical practice remains hindered by its limited ability to perform complex reasoning, utilize specialized tools, or provide transparent justifications. We introduce OphAgent, a multi-agent framework powered by large language models that mimics clinical workflows through dynamic diagnostic planning, knowledge base consultation, and multimodal image interpretation via structured inter-agent debate. Validated across four imaging modalities, 30+ conditions, and 11 diverse real-world datasets, OphAgent achieved 85% average accuracy in retinal diagnosis, surpassing current state-of-the-art multimodal large language models. In a large-scale global reader study with 311 ophthalmologists from 86 centres across 22 countries, spanning seven languages and settings from tertiary hospitals to resource-limited clinics, OphAgent improved diagnostic accuracy by up to 20% across 14 clinically significant tasks. Critically, these gains remained stable across geographic regions, experience levels, and languages, demonstrating robustness, fairness, and generalisability. OphAgent exemplifies an emerging paradigm where medical AI evolves from isolated prediction to collaborative clinical reasoning. By integrating reasoned clinical consultation with state-of-the-art diagnostic performance, the framework offers a scalable pathway toward more equitable, transparent, and trustworthy AI-enabled clinical decision support, with the potential to extend expert-level ophthalmic assessment across diverse healthcare settings.

Background Accurate estimation of the retinal age, defined as the age predicted from fundus photographs by a deep-learning model trained on chronological age, provides a non-invasive biomarker of biological ageing and disease risk. Methods In this study, we trained an ensemble multitask learning model that integrates fundus photographs with glycated haemoglobin using 50,595 quality-controlled fundus photographs from 27,214 disease-free adults and validated it on an independent set of 7288 additional images from disease-free adults. Model performance was evaluated using mean absolute error. Prediction uncertainty was quantified by calculating the standard deviation across ensemble predictions for each eye, and eyes were stratified based on this standard deviation. Results Here we show that the model achieves mean absolute errors of 2.78 years in internal validation and 3.39 years and 8.63 years in two external cohorts comprising 135 and 4992 eyes, respectively. Eyes with ensemble standard deviations below the median demonstrate improved age-prediction accuracy (mean absolute error: 2.46 years internally; 2.87 years in the primary external cohort). In a systemic disease cohort of 8467 individuals, the retinal age gap (predicted minus chronological age) is significantly higher in participants with diabetes, cardiac disease, or stroke after adjustment for age and sex, indicating older-appearing retinas and supporting the biological relevance of retinal age. Conclusions Retinal age derived from a single fundus photograph provides a scalable and clinically deployed biomarker of biological ageing. This approach may enable opportunistic screening for cardiometabolic and other ageing-related diseases in other routine ocular imaging workflows.

Objectives: Previously, we identified four open-angle glaucoma (OAG)-associated autoantibodies (anti-ETNK1, anti-VMAC, anti-NEXN, and anti-SUN1) using proteome-wide autoantibody screening by wet protein arrays. The objective of this exploratory study was to evaluate the diagnostic performance of these four glaucoma-associated autoantibodies using automated machine learning. Methods: Plasma samples from 119 patients with OAG and 35 patients with cataracts as controls were enrolled for the study. All machine-learning analyses were performed in Python 3.9.16 (GCC 11.2.0) using scikit-learn 1.2.2 and PyCaret 3.0.1. Variables included plasma levels of the autoantibodies, age, sex, and intra-ocular pressure (IOP). Probability calibration (Platt/sigmoid and isotonic) was assessed with reliability curves and Brier scores. Model explainability was examined with permutation importance, SHAP values, and an ablation analysis removing one autoantibody at a time. Results: The tuned random forest achieved an out-of-fold (OOF) area under the receiver-operating characteristic curve (ROC–AUC) of 0.852 (±0.040), an average precision (AP) of 0.950, and an F1 score of 0.865. Isotonic mapping improved agreement between predicted and empirical probabilities. Among these four autoantibodies, VMAC was the most important factor for the model’s prediction. Conclusions: A machine learning model using four autoantibodies from blood samples showed potential for diagnosing OAG.

Objectives: The objective of this study was to identify novel autoantibodies specific for open-angle glaucoma (OAG), including normal-tension glaucoma (NTG) and primary open-angle glaucoma (POAG), using proteome-wide autoantibody screening and to determine their utility for diagnosis. Methods: We conducted proteome-wide autoantibody screening by wet protein arrays. Autoantibody reactivity in the plasma of OAG patients (50 NTG and 69 POAG patients) was quantitatively analyzed and compared to that of controls (35 cataract patients). The area under the curve (AUC) of the receiver operating characteristic (ROC) and multivariate analyses were used to determine diagnostic potential in patients with OAG. Results: Based on differences in autoantibody titers and positivity rates, four autoantibodies against ETNK1, VMAC, NEXN, and SUN1 were selected as potential biomarkers to discriminate OAG and cataract. In discrimination between POAG and cataract, the AUCs of ETNK1 and VMAC were calculated to be 0.820 (95%CI: 0.733–0.907) and 0.889 (95%CI: 0.818–0.959), respectively. Furthermore, the combination of these four antibodies demonstrated diagnostic potential for OAG with an AUC of 0.828 (95%CI: 0.757–0.898) by multivariate analysis. Conclusions: Four new glaucoma-associated autoantibodies were identified in this study. The differences in autoantibody patterns in the plasma between glaucoma and cataract patients support their potential utility as biomarkers for glaucoma screening.

Glaucoma, a leading cause of blindness, requires accurate early detection. We present an AI-based Glaucoma Screening (AI-GS) network comprising six lightweight deep learning models (total size: 110 MB) that analyze fundus images to identify early structural signs such as optic disc cupping, hemorrhages, and nerve fiber layer defects. The segmentation of the optic cup and disc closely matches that of expert ophthalmologists. AI-GS achieved a sensitivity of 0.9352 (95% CI 0.9277–0.9435) at 95% specificity. In real-world testing, sensitivity dropped to 0.5652 (95% CI 0.5218–0.6058) at ~0.9376 specificity (95% CI 0.9174–0.9562) for the standalone binary glaucoma classification model, whereas the full AI-GS network maintained higher sensitivity (0.8053, 95% CI 0.7704–0.8382) with good specificity (0.9112, 95% CI 0.8887–0.9356). The sub-models in AI-GS, with enhanced capabilities in detecting early glaucoma-related structural changes, drive these improvements. With low computational demands and tunable detection parameters, AI-GS promises widespread glaucoma screening, portable device integration, and improved understanding of disease progression.