Research map / Vision-language and multimodal / Cross-modal learning

Risk Horizons: Structured Hypothesis Spaces for Longitudinal Clinical Prediction

NeurIPS 2026 · Vision-language and multimodal › Cross-modal learning · approach: Medical multimodal fusion · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Risk Horizons, which builds patient-specific structured candidate spaces for multi-modal next-visit EHR prediction by combining ICD/ATC hierarchies with lagged PMI cross-modal edges, embedding them in hyperbolic space, and retrieving candidates via directional risk cones, with optional constrained LLM reranking.

Problem

Next-visit prediction from longitudinal EHRs requires selecting plausible future diagnoses, procedures, and medications from a large, sparse, weakly structured event space; cross-modal and temporal dependencies are not specified by coding hierarchies.

Approach

Reframes next-visit prediction as constructing a patient-specific candidate hypothesis space (Risk Horizon) via hyperbolic embeddings of a graph combining coding hierarchies with lagged PMI edges, retrieved with directional tangent-space cones, and ranked geometrically with optional LLM reranking restricted to candidates.

Key results

• Risk Horizons (geometry-only) reaches 0.357 overall R@10 and 0.447 nDCG@10 on MIMIC-IV, versus 0.363 R@10 for RETAIN and 0.402 for Med-BERT.
• RH + Llama-3.3-70B-Instruct reranking reaches 0.425 R@10 and 0.516 nDCG@10 overall on MIMIC-IV, the best overall result.
• On eICU, RH + Llama-3.3-70B-Instruct achieves 0.401 R@10 and 0.433 nDCG@10, while RETAIN obtains 0.381 R@10 and 0.424 nDCG@10.

Surprising findings

• Removing lagged PMI stitching collapses R@10 from 0.357 to 0.005, showing deterministic hierarchies alone are insufficient.
• Euclidean RH paired with LLM reranking often does not improve and may reduce performance, suggesting downstream reasoning cannot repair a poor candidate space.
• Medication-only Oracle and Risk Horizon recall are high, but Med-BERT and RETAIN outperform the proposed geometry-only RH on medication nDCG@10 (RETAIN 0.723 vs RH 0.705).

Limitations

• Performance degrades on highly novel or externally driven events (e.g., trauma) weakly supported by history.
• LLM-only baselines perform poorly on frequent-event candidate sets, and LLM gains depend on the candidate set quality.
• Lagged PMI edges are associational, not causal, and depend on training-split statistics.

Experimental setup

Datasets: MIMIC-IV eICU

Benchmarks: MIMIC-IV eICU

Metrics: Recall@10 nDCG@10 MRR Candidate Recall Risk-Horizon Recall@K Oracle Recall@K Hierarchy Distance Ancestor Match@2

Baselines: Copy Last Visit RETAIN Med-BERT HALO Euclidean RAG Euclidean Risk Horizons Hyperbolic Linear Probe Hyperbolic NN Global

Builds on: RETAIN HALO Med-BERT GRAM

Closest papers by meaning

• Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data (ICLR 2025 · similarity 0.88)
• One Loss to Rule Them All: Marked Time-to-Event for Structured EHR Foundation Models (NeurIPS 2026 · similarity 0.88)
• Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval (ICLR 2025 · similarity 0.86)
• SAFER: A Calibrated Risk-Aware Multimodal Recommendation Model for Dynamic Treatment Regimes (ICML 2025 · similarity 0.85)
• History-Aware Conformal Prediction Sets for Censored Time-to-Event Outcomes (NeurIPS 2026 · similarity 0.85)
• Time-to-Event Pretraining for 3D Medical Imaging (ICLR 2025 · similarity 0.85)