A new preprint on arXiv proposes HypTI, a framework designed for AI-assisted dementia detection using both tabular clinical data and neuroimaging. The authors argue that existing multimodal fusion approaches typically combine features in Euclidean space, which may be a poor fit for the complex, hierarchical patterns found in dementia-related data.
HypTI instead uses hyperbolic geometry to fuse the two data types, aiming to preserve structural relationships that Euclidean space might flatten. The paper introduces this approach as a geometry-aware alternative, though full experimental details are not included in the abstract.
The work highlights a growing interest in moving beyond standard Euclidean assumptions for medical machine learning. As of the announcement, no comparison results or performance figures were provided, so the practical benefits of HypTI remain to be evaluated against existing methods.