Multimodal Classification of Keratoconus Progression Using Corneal Surface Geometry and Refractive Data: The Role of Peripheral Regions and Missing-Data Patterns

Early detection of ocular pathologies is crucial for improving treatment outcomes and preventing severe vision impairment. Keratoconus is a progressive corneal disorder characterized by thinning and irregular deformation of the cornea, which leads to visual distortion and may ultimately require corneal transplantation. Accurate detection and staging of keratoconus progression are therefore essential for effective clinical management. This study presents a multimodal framework for classifying keratoconus stages by integrating raw anterior and posterior corneal point cloud data acquired with Sirius topographer together with clinically derived refractive parameters. The proposed approach combines surface-based deep learning with traditional machine learning models for tabular clinical data, enabling joint exploitation of structural and optical information. Beyond classification performance, the study investigates the diagnostic role of peripheral corneal regions and acquisition-related missing-data patterns through dedicated analyses. The results indicate that peripheral measurements and structured missing-data distributions contain diagnostically relevant information that complements central corneal features. Overall, the proposed multimodal strategy improves keratoconus classification performance while providing new insights into the spatial and acquisition-related cues that contribute to disease staging, highlighting the potential of data-driven approaches to support clinical decision-making in keratoconus management.