Reproduce “MedFuse: Multiplicative Embedding Fusion For Irregular Clinical Time Series” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yi-Hsien Hsieh, Ta-Jung Chien, Chun-Kai Huang, Shao-Hua Sun, Che Lin
Researchers at National Taiwan University introduced MedFuse, a framework featuring MuFuse, a novel multiplicative embedding fusion module that captures value-dependent, nonlinear interactions in irregular clinical time series. MedFuse consistently established new state-of-the-art performance, achieving the highest AUPRC across three real-world clinical datasets, including ICU mortality prediction and chronic disease prognosis.
Key insights
- Multiplicative embedding fusion significantly enhances model expressiveness by enabling value-dependent, nonlinear modulation of feature identity embeddings, outperforming additive and concatenative approaches.
- Previous state-of-the-art methods like SCANE's value-scaling mechanism are identifiable as a specific instance of the more generalized MuFuse framework.
- Feature identity embeddings learned by MedFuse demonstrate transferability across clinical cohorts, indicating potential for pretraining on large datasets to improve performance on smaller target datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.