Papers of Note

When Does Multimodal Learning Help in Healthcare? A Benchmark on EHR and Chest X-Ray Fusion

Kejing Yin, Haizhou Xu, Wenfang Yao, Chen Liu, Zijie Chen, Yui Haang Cheung, William K. Cheung, Jing Qin

Introduces CareBench, a comprehensive benchmark evaluating multimodal fusion of EHR and chest X-rays across 14 fusion methods, 3 clinical tasks, and realistic missing-data scenarios. Reveals that fusion helps most for modality-distributed diseases, that modality balancing matters more than architectural complexity, and that multimodal models do not inherently improve algorithmic fairness.

mHC: Manifold-Constrained Hyper-Connections

Zhenda Xie, Yixuan Wei, Huanqi Cao, et al.

A framework that addresses training instability and memory overhead in Hyper-Connections architectures by projecting residual connection space onto specific manifolds. Demonstrates improved scalability for large-scale model training.

Emergent Hierarchical Reasoning in LLMs through Reinforcement Learning

Haozhe Wang, Qixin Xu, Che Liu, Junhong Wu, Fangzhen Lin, Wenhu Chen

Demonstrates how reinforcement learning enhances LLM reasoning through a two-phase hierarchy: initial low-level skill development followed by high-level strategic planning. Introduces HICRA (Hierarchy-Aware Credit Assignment) to optimize planning tokens and strengthen strategic reasoning.

Foundation Models for Medical Imaging: A Critical Review

An expert analysis of recent foundation model papers, examining their claims about generalization, data efficiency, and clinical applicability.

Algorithmic Bias in Healthcare: Key Papers and Lessons

A comprehensive review of seminal papers on bias in medical AI, from Obermeyer's cost prediction study to pulse oximetry disparities.

Real-World AI Deployment: What the Literature Reveals

Analysis of papers documenting successful (and failed) clinical AI deployments, extracting practical lessons for implementation.

External Validation in Medical AI: A Systematic Review

Why do most medical AI models fail external validation? An analysis of validation methodologies and common pitfalls.