Applications of Artificial Intelligence in the Diagnosis and Management of Keratitis and Conjunctivitis: A Review

Ocular surface diseases, such as conjunctivitis and keratitis, are an important global health concern because of their socioeconomic burden and potential for visual impairment.Early diagnosis and appropriate management are key to preventing complications, yet traditional diagnosis based on clinical examination and investigations is resource-intensive and varies among clinicians.Artificial intelligence (AI), particularly machine learning and deep learning, has shown promise in image-focused specialties such as ophthalmology.This narrative review summarizes published studies evaluating AI in the diagnosis and management of keratitis and conjunctivitis, including image-based diagnostic models, chatbots, natural language processing tools, and smartphone applications, along with future directions and ethical considerations.Convolutional neural networks applied to slit-lamp, anterior segment optical coherence tomography, in vivo confocal microscopy, and smartphone images showed high performance in detecting infectious keratitis, with areas under the curve above 0.96 in large multicenter datasets.Performance was lower for differentiating causative organisms, with accuracies of approximately 76% to 92% for bacterial, fungal, and Acanthamoeba keratitis and fungal subtypes.In allergic conjunctival disease, an AI pipeline achieved a diagnostic accuracy of 86.2%, compared with 60.0% for board-certified ophthalmologists.Chatbots correctly answered 51% to 68% of questions on corneal and conjunctival diseases, and incorrect AI suggestions significantly reduced the diagnostic accuracy of non-specialist ophthalmologists.Ensemble learning, mobile imaging applications, and chatbot-supported triage may extend tele-ophthalmology to under-resourced areas.However, most models rely on retrospective, single-center datasets, and concerns remain regarding generalizability, accountability, transparency, and clinician trust.Prospective multicenter validation and explainable, multimodal systems are needed before AI can be routinely adopted as a supportive tool in clinical practice.