GraphDiffMed uses noise-aware attention and drug graphs for safer medication recommendations
Dual-scale differential attention filters spurious signals across patient visits...
Recommending safe, effective medication combinations from electronic health records (EHRs) is a core clinical AI challenge. Patient trajectories are long, noisy, and heterogeneous. Existing methods typically excel at either temporal modeling across visits or integrating pharmacological knowledge (e.g., drug-drug interactions), but rarely achieve both while robustly suppressing noise. Saxena and Shibata introduce GraphDiffMed, a knowledge-constrained framework built on dual-scale Differential Attention v2. It applies differential attention at both intra-visit and inter-visit levels to filter spurious signals within encounters and across longitudinal history, while pharmacological constraints are incorporated during learning.
Experiments on MIMIC-III and ablation studies show GraphDiffMed consistently improves recommendation quality and ranking over strong baselines, achieving a more favorable safety-performance balance. Notably, the strongest configuration uses only demographic auxiliary features. By combining noise-aware attention with pharmacological priors, GraphDiffMed yields more reliable and clinically meaningful medication recommendations. The authors open-source their code, enabling further research and potential integration into clinical decision support systems.
- Uses Differential Attention v2 at both intra-visit and inter-visit scales to filter noise from EHR sequences
- Incorporates pharmacological graph priors (drug-drug interactions) as learning constraints
- Outperforms strong baselines on MIMIC-III with improved safety-performance balance
Why It Matters
More reliable AI medication recommendations could reduce adverse drug events in clinical settings.