GUIDE: A Responsible Multimodal Approach for Enhanced Glaucoma Risk Modeling and Patient Trajectory Analysis
Keywords: Multimodal AI, Responsible AI, Interpretable Machine Learning, Contextual Transparency, Hierarchical Representation Learning
TL;DR: GUIDE presents a conceptual framework for responsible, interpretable, and multimodal AI in glaucoma management, addressing critical challenges in healthcare AI while building reliable and ethical foundational models across diverse modalities.
Abstract: Glaucoma is a irreversible vision loss that disproportionately affects marginalized communities. Current diagnostic and management strategies often fail to account for individualized risks, leading to suboptimal patient outcomes and exacerbating health disparities. Here, we present GUIDE (Glaucoma Understanding and Integrated Data Evaluation), a conceptual framework for explainable multimodal AI framework that integrates diverse data sources—including clinical measurements, imaging data, unstructured electronic health records (EHR), and social determinants of health (SDOH)—to create a comprehensive and personalized view of glaucoma risk and progression. Our approach focuses on developing clinically interpretable, expert-tunable hierarchical fusion models that address key issues such as fairness, transparency, and robustness, aligning with responsible AI principles. By disentangling the embedding space using clinical supervision at each stage of modality fusion, we prevent model hallucinations and ensure that the embeddings can be decoded into physician-understandable clinical concepts. We also implement contextual transparency by engaging stakeholders and tailoring transparency measures according to the NIST’s Contextual Transparency Playbook. Our framework handles data quality issues through pre-training strategies and hierarchical data fusion, and considers modalities with varying costs to optimize resource utilization. We demonstrate how GUIDE provides a comprehensive understanding of glaucoma progression, facilitates more accurate risk stratification, and enables personalized treatment plans.
Submission Number: 44
Loading