HypTI: Hyperbolic Geometry-Aware Multimodal Fusion for Improved Dementia Detection Accuracy

Title

HypTI: Hyperbolic Geometry-Aware Multimodal Fusion for Improved Dementia Detection Accuracy

Meta Description

This paper proposes the HypTI framework, the first to introduce hyperbolic geometry into multimodal fusion of clinical tabular data and MRI, modeling evaluation hierarchies via Lorentz prototypes and capturing anatomical patterns through ROI-guided multi-scale encoding, with dementia detection performance validated on the OASIS-3 dataset.

Article Content

In early dementia detection, multimodal approaches that fuse clinical tabular data with MRI imaging are the mainstream line of thought, but these methods generally perform feature fusion in Euclidean space. A new paper proposes the HypTI framework, which introduces hyperbolic geometry into joint tabular-imaging modeling, attempting to more naturally characterize the hierarchical structure of clinical assessments and the multi-scale anatomical patterns of neuroimaging.

Why is Euclidean space insufficient?

Existing dementia detection models typically concatenate clinical scales (such as MMSE, CDR) with MRI features and feed them into a classifier. The problem is that clinical variables often have hierarchical relationships—a cognitive domain score is aggregated from multiple sub-items; brain region structures in MRI also exhibit multi-scale nested patterns. The flat metric of Euclidean space struggles to efficiently express such tree-like or hierarchical structures, and the model requires more parameters and capacity to "approximate" regularities that could be directly characterized by geometric priors.

The paper points out that hyperbolic space, due to its exponential expansion property, is naturally suited to modeling hierarchies and multi-scale structures. Recent studies have validated this advantage on images and graph data, but in the common combination of "tabular + MRI" for dementia detection, hyperbolic representation learning remains a blank. HypTI aims precisely at this gap.

Hyperbolic geometry and negative curvature: Hyperbolic space is a class of constant negative curvature space whose "volume" grows exponentially with radius, thus enabling low-dimensional embeddings to represent exponentially large hierarchical structures (such as trees). Euclidean space volume grows only polynomially, requiring high dimensionality and suffering large distortion when representing tree-structured data. Common models include the Poincaré disk, the upper half-plane, and the Lorentz model (hyperboloid model). The Lorentz model has better numerical stability and is suitable for gradient optimization, and has been widely used in hierarchical representation learning in recent years.

HypTI's three-layer design

The overall idea of HypTI is to uniformly handle three types of information in Lorentz hyperbolic space: clinical tabular data, MRI imaging, and cross-modal interaction.

Clinical tabular data: learnable Lorentz prototypes

For tabular variables, HypTI learns Lorentz prototype representations for each assessment item, mapping discrete or continuous clinical scores to points in hyperbolic space. Hierarchical relationships are naturally reflected through the positions of prototypes in hyperbolic space—parent nodes near the origin, child nodes extending outward—without the need to explicitly construct a tree structure.

Lorentz prototypes and hyperbolic distance: In the Lorentz model, points lie on a hyperboloid, and distances between points are defined by the hyperbolic metric rather than Euclidean distance. Prototype learning initializes a hyperbolic point for each clinical assessment item and measures sample-prototype similarity via hyperbolic distance. The hierarchical structure of parent nodes near the origin and child nodes extending outward emerges naturally through optimization, without predefining a tree. However, hyperbolic distance computation involves inverse hyperbolic functions, requiring attention to numerical stability and initialization strategies.

MRI: ROI-guided multi-scale encoding

The MRI branch employs region-of-interest (ROI)-guided multi-scale representation learning. Different brain regions are encoded at different spatial scales, and distance relationships in Lorentz space correspond to hierarchical patterns of anatomical structures. Compared to directly flattening voxel features, this design preserves organizational information among regions.

ROI and multi-scale brain region encoding: ROI (region of interest) refers to brain regions delineated according to prior atlases, such as the hippocampus and entorhinal cortex, structures closely associated with dementia. Multi-scale encoding refers to extracting features at different spatial resolutions or brain region granularities—for example, global brain regions, sub-regions, and even voxel-level local patterns—then uniformly mapping them to hyperbolic space. The hierarchical nested relationships of anatomical structures (brain → lobe → gyrus → nucleus) align with the tree-like prior of hyperbolic geometry, so ROI-guided encoding can preserve inter-regional organizational information rather than simply flattening voxels.

Adaptive fusion: generating modulation coefficients for each subject

The embeddings of the two modalities are not simply concatenated. HypTI designs an adaptive fusion module that generates subject-specific scale and shift coefficients for each subject, dynamically modulating and integrating multimodal embeddings. This "conditional" fusion allows the model to adjust the contribution weights of each modality according to individual differences, ultimately outputting a patient-level joint tabular-imaging representation for dementia detection.

Conditional feature modulation (FiLM): HypTI generates scale and shift coefficients for each subject, applying an affine transformation to multimodal embeddings; this mechanism is similar in spirit to Feature-wise Linear Modulation (FiLM). FiLM generates modulation parameters for each sample through a conditioning network, enabling the model to dynamically adjust the contribution of each modality and achieve individualized fusion. In hyperbolic space, modulation must be performed in the tangent space or via exponential/logarithmic mappings to preserve geometric structure, making implementation more complex than in Euclidean space.

Experimental results and limitations

The paper conducts experiments on the OASIS-3 dataset, and HypTI achieves strong performance compared to representative methods. OASIS-3 is a public longitudinal neuroimaging dataset containing clinical assessments and MRI, suitable for validating tabular-imaging fusion tasks.

It should be viewed objectively that the paper abstract does not disclose specific metric values or ablation details, and the magnitude of the gain brought by hyperbolic space still needs to be corroborated by the full paper's data. In addition, the initialization strategy of Lorentz prototypes, the degree of standardization of ROI definitions, and the model's generalization ability on more heterogeneous cohorts are all issues worthy of attention in the future.

Implications for multimodal medical AI

The value of HypTI goes beyond dementia detection. It suggests a more general direction: when multimodal data contain hierarchical or multi-scale structures, the geometric space in which fusion occurs is itself a designable variable. The combination of clinical tabular data and imaging widely exists in scenarios such as Alzheimer's disease and Parkinson's disease, and hyperbolic geometry-aware fusion ideas have transfer potential.

Of course, the introduction of geometric priors also brings new challenges in training stability and interpretability. Hyperbolic operations are sensitive to numerical precision, and more thorough engineering validation is needed before large-scale deployment. HypTI has taken an exploratory step; whether it becomes a standard practice depends on subsequent reproduction results on more datasets and tasks.

One-line Summary

HypTI introduces hyperbolic geometry into tabular-MRI fusion, using Lorentz space to model hierarchical structures in dementia detection.

Paragraph Summary

The paper proposes the HypTI framework to address the fusion of clinical tabular data and MRI imaging in early dementia detection, pointing out that existing methods fuse in Euclidean space and struggle to express the hierarchical structure of clinical assessments and the multi-scale anatomical patterns of MRI. HypTI uniformly models in Lorentz hyperbolic space: clinical tabular data uses learnable Lorentz prototypes to capture hierarchical relationships, MRI uses ROI-guided multi-scale encoding to preserve regional organizational information, and an adaptive fusion module is designed to generate scale and shift coefficients for each subject to dynamically modulate multimodal embeddings. It achieves strong performance compared to representative methods on the OASIS-3 dataset, but specific metrics and ablation details are not disclosed in the abstract, and the magnitude of hyperbolic gains, initialization strategies, ROI standardization, and generalization ability still require further validation. The article believes this approach can be transferred to other multimodal medical scenarios but must address challenges in hyperbolic training stability and interpretability.