PaperlayerView exact revision on arXiv

Source-linked AI summary

Conditional Generation Net for Medication Recommendation

Rui Wu, Zhaopeng Qiu, Jiacheng Jiang, Guilin Qi, Xian Wu

arXiv:2202.06588v2cs.LG

TL;DR

Medication recommendation is challenging for patients with multiple diseases because clinicians must treat diagnosed conditions while avoiding harmful drug-drug interactions. COGNet generates medications sequentially, choosing whether to copy historical recommendations or predict new medicines, and experiments on MIMIC-III report that it outperforms existing methods.

  • Problem

    Medication recommendation must select medicines for multiple diagnosed diseases while avoiding harmful drug-drug interactions, a task that is difficult to conduct manually in complicated cases.

  • Method

    COGNet is an encoder-decoder medication recommendation network that uses a copy-or-predict mechanism to copy historical medicines or predict new ones.

  • Results

    COGNet outperforms existing medication recommendation methods on the publicly available MIMIC-III dataset.

  • Takeaways & Limitations

    COGNet can leverage historical medication information and multiple past visits in medication recommendation.

  • Takeaways & Limitations

    The paper notes that rule-based methods require substantial clinician effort and lack generalization.

Abstract

from arXiv · show

Medication recommendation targets to provide a proper set of medicines according to patients' diagnoses, which is a critical task in clinics. Currently, the recommendation is manually conducted by doctors. However, for complicated cases, like patients with multiple diseases at the same time, it's difficult to propose a considerate recommendation even for experienced doctors. This urges the emergence of automatic medication recommendation which can help treat the diagnosed diseases without causing harmful drug-drug interactions.Due to the clinical value, medication recommendation has attracted growing research interests.Existing works mainly formulate medication recommendation as a multi-label classification task to predict the set of medicines. In this paper, we propose the Conditional Generation Net (COGNet) which introduces a novel copy-or-predict mechanism to generate the set of medicines. Given a patient, the proposed model first retrieves his or her historical diagnoses and medication recommendations and mines their relationship with current diagnoses. Then in predicting each medicine, the proposed model decides whether to copy a medicine from previous recommendations or to predict a new one. This process is quite similar to the decision process of human doctors. We validate the proposed model on the public MIMIC data set, and the experimental results show that the proposed model can outperform state-of-the-art approaches.

1 INTRODUCTION

Medication recommendation is difficult for patients with multiple diseases because doctors must select treatments while avoiding harmful drug-drug interactions. COGNet addresses this challenge by generating medications sequentially through a copy-or-predict mechanism that uses historical recommendations, and experiments on MIMIC-III show effectiveness.

  • Medication recommendation requires selecting medicines for diagnosed diseases while avoiding harmful drug-drug interactions.
  • Existing instance-based models use current diagnoses and procedures, whereas longitudinal models incorporate patients’ historical information.
  • Historical recommendations are closely related to current medications, especially for patients with chronic diseases who may continue using the same medicines.
  • COGNet generates medications sequentially with a copy-or-predict mechanism that either copies historical medicines or predicts new ones.
  • COGNet uses hierarchical selection to choose reusable medicines from both medication-level and visit-level perspectives.
  • Comprehensive experiments on the public MIMIC-III dataset demonstrate the effectiveness of COGNet.

2 PROBLEM FORMULATION

The formulation represents each patient as a longitudinal sequence of visits containing diagnoses, procedures, and medications. COGNet uses current diagnoses, procedures, historical visits, and EHR/DDI graphs to recommend the medication combination for the current visit.

  • 2.1 Electrical Health Records (EHR): A longitudinal EHR represents each patient as a sequence of multivariate observations across hospital visits.
  • 2.1 Electrical Health Records (EHR): For a single patient, each visit is represented by diagnoses, procedures, and medications drawn from their respective global sets.
  • 2.2 EHR&DDI Graph: The EHR graph connects medications that have appeared together, while the DDI graph records known mutually interacting medication pairs.
  • 2.2 EHR&DDI Graph: The graph adjacency matrices encode medication co-occurrence in the EHR graph and pairwise drug-drug interactions in the DDI graph.
  • 2.3 Medication Recommendation Problem: Given current diagnoses, procedures, historical visits, and the EHR and DDI graphs, COGNet recommends the current medication combination.
  • 2.3 Medication Recommendation Problem: The formulation’s notation defines the sets and visit variables used to describe diagnoses, procedures, medications, and longitudinal patient records.

3 FRAMEWORK

COGNet is an encoder-decoder framework that generates medications sequentially from current health conditions while incorporating historical visits. Its copy module selects reusable medicines from similar past visits, while the decoder models medication relations and uncovered diseases.

  • Overall Architecture: COGNet encodes current and historical medical codes, then generates the medication combination one medicine at a time.The decoder combines diagnoses, procedures, and previously generated medications when recommending the next medicine.
  • Basic Model: The basic model recommends medications from the patient’s current health conditions using diagnosis, procedure, medication-graph, and medication-combination modules.Its encoder-decoder generation process does not use historical visits.
  • Basic Model: The diagnosis encoder uses multi-head self-attention to capture relations among diagnoses within each visit, with residual connections and layer normalization.The procedure encoder uses the same network structure with different parameters.
  • Medication Graph Encoder: The medication graph encoder models medication co-occurrence and DDI relations with separate two-layer GCNs, then fuses their representations.Co-occurrence information supports recommendations based on partially generated medications, while DDI modeling supports avoiding conflicts.
  • Medication Combination Decoder: During decoding, self-attention captures interactions among generated medications, while aligned health-condition representations guide recommendations toward uncovered diseases.The decoder predicts the next medication from the partially generated combination and current visit conditions.
  • Copy Module: The copy module compares current and historical health conditions, selects a similar past visit, and then selects reusable medications from that visit.This hierarchical selection operates at visit and medication levels during each decoding step.

4 EXPERIMENTS

Experiments on MIMIC-III compare COGNet with established baselines, assess historical-visit effects, ablate its components, and illustrate its copy-or-predict behavior. COGNet generally outperforms baselines and ablations while copying reusable medicines and generating new ones.

  • Overall Comparison: COGNet outperforms all baselines on Jaccard, F1, and PRAUC, while SafeDrug achieves a lower DDI rate through drug-molecule information.COGNet has a DDI rate similar to MIMIC-III’s average DDI of 0.08379.
  • Overall Comparison: COGNet achieves relatively better performance as the number of patient visits increases, compared with SafeDrug and MICRON.The analysis uses the first five visits because most MIMIC patients have fewer than five visits.
  • Ablation Study: Removing the copy module significantly worsens the basic model, indicating that replicating historical drugs improves medication recommendation.The ablation also evaluates visit-level selection, graphs, diagnoses, procedures, and beam search.
  • Ablation Study: Removing diagnoses or procedures produces poor results, while removing graphs or beam search also reduces performance; the complete COGNet performs best.These results support the contribution of each evaluated component to the final recommendation.
  • Case Study: In the case study, COGNet assigns high copy probabilities to reusable medicines and generates new drugs for newly diagnosed conditions.The visualization suggests that the copy mechanism supports both historical-drug reuse and new-drug generation, while making the process interpretable.

5 RELATED WORK

Related work spans rule-based, instance-based, longitudinal, and graph-enhanced approaches to medication recommendation. These methods differ in whether they rely on clinician protocols, current-visit information, patient history, or structured drug representations.

  • Rule-based Methods: Rule-based methods rely on human-designed recommendation protocols but require substantial clinician effort and lack generalization.
  • Instance-based Methods: Instance-based methods use only the current visit’s information and therefore ignore the patient’s historical visits.
  • Longitudinal Methods: Longitudinal methods use historical patient information and model sequential dependencies between visits, often with RNNs or memory networks.Some approaches additionally incorporate drug-molecule information to improve medication representations.
  • Graph Neural Networks: Graph neural networks learn node representations by aggregating neighboring features, and medication-recommendation studies use them for patient or drug representations.One cited approach applies GNNs to drug-molecule structures.

6 CONCLUSION

The conclusion presents COGNet as a medication-recommendation model that leverages historical information at the medication level through copy-or-predict generation. Experiments show improved performance over existing methods, effective use of multiple visits, and contributions from its individual modules.

  • Conclusion: COGNet uses an encoder-decoder framework and a copy-or-predict mechanism to determine whether historical medications remain relevant.The model is designed to leverage historical information from a medication-level perspective.
  • Conclusion: Experiments on MIMIC-III show that COGNet outperforms existing medication-recommendation methods.
  • Conclusion: Visit-count analysis indicates that COGNet effectively incorporates information from multiple past visits, while ablations support the effectiveness of each module.

A.1 Dataset Processing

The study processes MIMIC-III records by merging admission-linked clinical tables, retaining patients with repeated visits, limiting medications and DDI types, and ordering medications by training-set frequency.

  • Data Sources and Merging: The source data come from MIMIC-III admission, diagnosis, procedure, and prescription tables merged through admission and subject identifiers.In this paper, an admission corresponds to a visit.
  • Filtering: Patients with at least two visits are retained, and the 300 most frequent medications are kept to improve training speed and facilitate analysis.
  • DDI Processing: The processing extracts the top 40 severity DDI types from TWOSIDES using ATC Third Level codes.
  • Medication Ordering: Medication frequencies are counted on the training set and used to sort medications in ascending order for all patients.

A.2 Implementation Details

The method is implemented with specified software, hardware, optimization, and generation settings, with hyperparameters selected on the validation set.

  • The implementation uses PyTorch 1.9.0 with Python 3.9.6 on an Intel Xeon 8255C machine with 315G RAM and 8 NVIDIA Tesla V100 GPUs.
  • Validation-set selection produced dimension size s = 64, 4 beam search states, and maximum generation length 45.
  • Models are trained with Adam at learning rate 1×10^-4, batch size 16, and 50 epochs.
  • The random seed is fixed at 1203 for PyTorch to support reproducibility.

A.3 Metrics

The experiments evaluate medication-set recommendations using overlap, retrieval-quality, and drug-interaction metrics, with PRAUC adapted to sequential generation.

  • Jaccard measures overlap between predicted and target medication sets, while F1 is the harmonic mean of precision and recall.
  • PRAUC assigns each medication a probability derived from its generation step or averaged across steps when it is not recommended.
  • DDI measures interactions among recommended medications using the DDI graph adjacency matrix.

B ADDITIONAL EXPERIMENTS

The additional experiment studies how medication-label ordering affects COGNet across training epochs, finding rare-first ordering strongest among the tested heuristics.

  • The experiment evaluates four ordering heuristics: rare first, frequent first, early first, and late first.
  • Rare first ranks medications by training frequency and places less frequent medications earlier.
  • Rare first outperforms the alternatives, while frequent first converges faster but performs poorly because popular medications dominate the recommendations.
  • Early-first and late-first results are weak, indicating that chronological ordering is not effective for this task; random disruption performs worse than the tested heuristics.
Loading 2202.06588v2…