HICL-TBI: Hyperbolic In-Context Learning for Traumatic Brain Injury Outcome Prediction

Tabular foundation models like TabPFN excel at training-free In-Context Learning (ICL), yet standard retrieval strategies rely on Euclidean geometry, failing to capture the hierarchical progressions inherent in clinical data. In this paper, we introduce HICL-TBI, a novel Hyperbolic ICL framework that integrates Poincaré Ball geometry to resolve this representational bottleneck. To robustly handle real-world data imperfections, we propose Hyperbolic Fréchet Imputation to reconstruct missing values without flat-space distortion, and Density-Aware Hyperbolic Retrieval to dynamically adapt the context size based on local manifold density. By mapping mild cases near the origin and exponentially separating severe outliers, our approach effectively mitigates class imbalance and the Euclidean crowding problem. Extensive experiments on three real-world Traumatic Brain Injury (TBI) datasets demonstrate that HICL-TBI outperforms Euclidean baselines in fine-grained classification and small-data regimes, all while maintaining a strictly training-free pipeline.