Source-linked AI summary
GAMENet: Graph Augmented MEmory Networks for Recommending Medication Combination
Junyuan Shang, Cao Xiao, Tengfei Ma, Hongyan Li, Jimeng Sun
TL;DR
Medication recommendation methods may fail to personalize recommendations from patient history or incorporate DDI knowledge, which is important for avoiding harmful interactions. GAMENet combines longitudinal patient representations with graph-augmented memory integrating EHR and DDI graphs, and it outperformed baselines while reducing DDI rates by 3.60%.
Problem
Existing medication recommendation approaches may omit longitudinal patient history or DDI knowledge, although DDIs can cause health deterioration or death.
Method
GAMENet uses Dual-RNN patient queries and graph-augmented memory integrating EHR drug-usage and DDI graphs for end-to-end medication-combination recommendation.
Results
GAMENet outperformed all baselines in effectiveness measures and achieved a 3.60% DDI rate reduction from existing EHR data.
Takeaways & Limitations
The results support GAMENet as an effective and safer approach to medication-combination recommendation on real-world EHR data.
Takeaways & Limitations
GAMENet may recommend medication pairs whose combined use could cause allergic bronchitis because of a trade-off between effectiveness and safety.
Abstract
from arXiv · showhide
Recent progress in deep learning is revolutionizing the healthcare domain including providing solutions to medication recommendations, especially recommending medication combination for patients with complex health conditions. Existing approaches either do not customize based on patient health history, or ignore existing knowledge on drug-drug interactions (DDI) that might lead to adverse outcomes. To fill this gap, we propose the Graph Augmented Memory Networks (GAMENet), which integrates the drug-drug interactions knowledge graph by a memory module implemented as a graph convolutional networks, and models longitudinal patient records as the query. It is trained end-to-end to provide safe and personalized recommendation of medication combination. We demonstrate the effectiveness and safety of GAMENet by comparing with several state-of-the-art methods on real EHR data. GAMENet outperformed all baselines in all effectiveness measures, and also achieved 3.60% DDI rate reduction from existing EHR data.
Introduction
GAMENet addresses medication recommendation gaps by combining longitudinal patient history with DDI knowledge to produce effective and safe medication combinations. It uses graph-augmented memory and patient representations, outperforming baselines and reducing DDI rates.
- Existing medication recommendation methods may omit patient history or drug-drug interaction knowledge, despite DDIs causing health deterioration or death.
- GAMENet jointly models longitudinal patient records and a DDI knowledge base for effective and safe medication-combination recommendation.
- The model combines Dual-RNN patient queries with a graph-augmented memory module integrating EHR drug usage and DDI graphs.
- A combined loss balances multi-label prediction for effectiveness against DDI loss for safety.
- 3.60% DDI rate reduction was achieved relative to existing EHR data, while GAMENet outperformed all baselines in effectiveness measures.
Related Works
Prior work includes memory and graph neural networks, alongside instance-based and longitudinal medication recommendation methods. Longitudinal approaches use temporal patient records, but the cited methods do not establish the integrated GAMENet design.
- Memory-augmented neural networks use external memory to help deep networks remember and store information, with early successes in NLP.
- Graph convolutional networks learn node embeddings by transforming weighted sums of neighboring node features through simultaneous propagation.
- Instance-based medication recommendation focuses on current health conditions, exemplified by Leap’s attention-based sequence-to-sequence model.
- Longitudinal methods leverage temporal clinical dependencies; RETAIN identifies influential visits and variables, while DMNC uses memory with recurrent encoder-decoder networks.
Method
GAMENet represents patient records and medication relationships as graphs, learns longitudinal patient queries with Dual-RNNs, and retrieves information from graph-augmented and dynamic memory to recommend medications.
- Patient records: Each patient is represented as a sequence of multivariate observations across visits, with each visit containing diagnosis and medication codes.
- Patient records: Medical codes are represented as multi-hot vectors, where each position indicates whether a code occurs at the visit.
- EHR and DDI graphs: The EHR graph encodes known medication combinations, while the DDI graph encodes pairwise drug-drug interactions.
- Graph-augmented memory: GAMENet defines the current patient representation as a query generated from longitudinal records and uses it to read from a graph-augmented memory bank and dynamic history memory.
- Recommendation: The query and memory outputs are concatenated for recommendation, while training optimizes a combined loss.
The GAMENet
GAMENet combines longitudinal EHR representations with graph-augmented memory to personalize medication combinations while incorporating drug-usage and DDI knowledge. Its differentiable memory reads patient history and graph embeddings, and training balances recommendation accuracy with DDI control.
- Architecture: GAMENet comprises medical embedding, patient representation, and graph augmented memory modules for medication recommendation.The patient representation uses Dual-RNNs, while the memory module integrates EHR and DDI graph information.
- Patient representation: Dual-RNNs separately encode diagnosis and procedure histories to learn patient representations from multimodal EHR data.The resulting hidden states form the patient health state and memory query.
- Graph augmented memory: The memory module stores EHR and DDI graph representations in Memory Bank and patient histories as key-value pairs in Dynamic Memory.Dynamic Memory supports retrieving similar historical patient representations and associated medication sets over time.
- Graph augmented memory: Two-layer GCNs learn embeddings from medication-usage and DDI graphs, which are fused into the Memory Bank representation.The weighting variable β controls fusion of the two graph embeddings.
- Memory reading: Attention-based reading retrieves graph and historical information for the patient query, producing memory outputs used for multi-label medication prediction.The reading procedure is differentiable, enabling end-to-end backpropagation.
- Training and inference: Training combines binary cross entropy, multi-label margin, and DDI losses to balance recommendation accuracy and medication safety.The DDI loss increases when predicted medication pairs have high interaction probability, while simulated-annealing-style weighting adjusts the balance during training.
Experiments
GAMENet was evaluated on MIMIC-III EHR data using medication-accuracy and DDI-safety measures against established baselines. It achieved the strongest reported accuracy and improved DDI-related outcomes, while a case study illustrated both its strengths and safety trade-offs.
- Safety results: 3.60% DDI rate reduction was achieved relative to existing EHR data.DDI Rate measures the percentage of medication recommendations containing drug-drug interactions.
- Overall results: GAMENet achieved the highest Jaccard, PR-AUC, and F1 scores among all compared baselines.These metrics evaluate medication recommendation accuracy.
- Baseline comparison: About 4% lower Jaccard and F1 scores were reported for Nearest and LR compared with GAMENet.Their DDI rates were also very close to the base DDI rate in the EHR data.
- Baseline comparison: RETAIN and DMNC outperformed instance-based Leap on accuracy, supporting the importance of temporal information in patients’ past EHRs.DMNC’s larger recommended medication sets may contribute to its high DDI rate.
- Ablation comparison: Adding DDI knowledge improved both performance and DDI rate over GAMENet without DDI knowledge, with statistical significance after ten runs.The comparison used GAMENet and its variant GAMENet (w/o DDI).
- Case study: In a two-visit case study, GAMENet produced 11 and 13 correct medications, missed 4 and 1, and still recommended one medication pair associated with allergic bronchitis.The example demonstrates a trade-off between effectiveness and safety when that pair also appeared in both ground-truth medication sets.
Conclusion
GAMENet recommends medication combinations using memory networks augmented with drug usage and DDI graphs, while incorporating patient history. On real-world EHR data, it outperformed all baselines in effectiveness measures and reduced the DDI rate by 3.60%.
- GAMENet uses memory networks augmented with integrated drug usage and DDI graphs, plus dynamic memory based on patient history.
- GAMENet outperformed all baselines in effectiveness measures on real-world EHR data.
- 3.60% DDI rate reduction was achieved relative to existing EHR data.