GraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation

2026-01-20 · Source: cs.LG updates on arXiv.org · Field: Health & Wellbeing — Clinical Care & Medical Practice, Pharmaceuticals & Biotechnology, Medical Devices & Health Technology · Depth: Expert, extended

Summary

GraphDiffMed is a novel knowledge-constrained framework designed for recommending safe and effective medication combinations from electronic health records (EHRs). It employs dual-scale Differential Attention v2 (DiffAttn_v2) at both intra-visit and inter-visit levels to effectively filter spurious signals within patient encounters and across longitudinal history. The framework integrates pharmacological constraints directly into the learning process. Evaluated on the MIMIC-III dataset, GraphDiffMed consistently outperforms strong baselines in recommendation quality and ranking, while also achieving a more favorable safety performance balance. Ablation studies reveal that the most effective configuration utilizes only demographic auxiliary features (gender and age). The authors open-source their code at https://github.com/saxenakrati09/GraphDiffMed, highlighting that combining noise-aware attention with pharmacological constraints yields more reliable and clinically meaningful medication recommendations.

Key takeaway

For Machine Learning Engineers developing clinical decision support systems, you should consider integrating noise-aware attention mechanisms and pharmacological graph priors into your medication recommendation models. GraphDiffMed demonstrates that dual-scale Differential Attention v2, combined with knowledge constraints, significantly improves recommendation quality and safety balance on MIMIC-III. Prioritize stable demographic features like age and gender as auxiliary inputs over high-dimensional, noisy lab events to achieve more reliable performance and a better quality-safety trade-off in your deployments.

Key insights

Combining dual-scale differential attention with pharmacological graph priors significantly enhances medication recommendation safety and quality.

Principles

• Dual-scale differential attention effectively suppresses noise in EHR data.
• Pharmacological graph priors can shape attention for improved safety-performance.
• Clean demographic features outperform noisy lab data as auxiliary inputs.

Method

GraphDiffMed employs multi-modal embeddings, graph-biased differential cross-attention for intra-visit context, and graph-biased differential attention for inter-visit history, followed by a causal review module and a multi-term objective with annealed DDI regularization.

In practice

• Apply dual-scale differential attention to filter EHR noise.
• Integrate DDI graphs as attention biases for safety.
• Prioritize demographics over noisy lab data for auxiliary features.

Topics

• Medication Recommendation
• Differential Attention
• Pharmacological Knowledge Graphs
• Drug-Drug Interactions
• Electronic Health Records
• Clinical Decision Support

Code references

• saxenakrati09/GraphDiffMed

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• OOD-GraphLLM: Graph Large Language Model for Out-of-Distribution Generalized Drug Synergy Prediction — OOD-GraphLLM uses a GraphLLM to predict drug synergy under out-of-distribution conditions by integrating molecular graph and semantic data.
• Case-Aware Medical Image Classification with Multimodal Knowledge Graphs and Reliability-Guided Refinement — Medical image classification improves by integrating multimodal knowledge graphs from similar cases with reliability-guided refinement.
• Gated Multi-Graph Fusion via Graph Attention Networks for Alzheimer's Disease Detection — Gated Multi-Graph Fusion via Graph Attention Networks improves Alzheimer's Disease detection from speech by integrating diverse linguistic features.
• AI scans 400,000 Reddit posts and finds hidden Ozempic side effects — AI-driven social media analysis can rapidly identify underreported drug side effects missed by traditional clinical trials.
• A Triple-Modal Contrastive Learning Framework with Sequence, Graph, and 3D Features for Drug-Target Interaction Prediction — TriMod-DTI uses triple-modal contrastive learning with 1D, 2D, and 3D features for superior drug-target interaction prediction.
• SPOKE: A massive biomedical knowledge graph for precision health and drug discovery — SPOKE integrates vast biomedical data into a knowledge graph for drug discovery and precision medicine.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.