AngleVLM-IoT: A Transfer-Learned Vision–Language Model for Automatic Anterior Chamber Angle Recognition in Connected Eye-Care Ecosystems

Accurately measuring the anterior chamber angle (ACA) is vital to identifying and treating angle-closure glaucoma, one of the top five causes of irreversible blindness globally. Current screening methods, mainly derived from imaging-based technologies, are constrained by reliance on expensive hardware equipment and highly educated personnel. This has a negative impact on the accessibility and affordability of these methods as a means of broad-scale outreach to underserved communities. In this article, we introduce AngleVLM-IoT, a new-sense vision-language artificial intelligence model that can automatically capture information about the ACA using consumer-level ophthalmic devices within the Internet of Things (IoT) ecosystem for eye health. We describe our two-part approach to building AngleVLM -IoT using transfer learning, starting with large training sets of multimodal biomedical images available via public imagers, and then further improving AngleVLM-IoT’s accuracy by using anterior segment optical coherence tomography (AS-OCT) images coupled with clinical data to fine-tune AngleVLM-IoT to become more effective at diagnosing using the anterior segment of the eye. To overcome the technical limitations of hosting our solution on IoT edge devices, we have successfully reduced the model’s total size by 84.2% via real-time dynamic knowledge distillation and mixed-precision quantization techniques-while still retaining all its diagnostic capabilities. Moreover, we have implemented a federated learning protocol within the AngleVLM-IoT model so that clinicians can train models on private patient data stored at various distributed IoT nodes located throughout a wide network of connected healthcare providers. Finally, after training AngleVLM-IoT using different patient populations, we found that the accuracy loss for previously unseen patient populations is no greater than 2.8%. We conclude that AngleVLM-IoT represents both a clinically relevant and an economically viable solution for rapidly performing assessments of ACA and enables widespread access to scalable methods to screen for angle-closure glaucoma among patients receiving routine care from consumer-oriented eye-care providers who lack access to gold-standard methods for screening.