# KnowAugNet: Multi-Source Medical Knowledge Augmented Medication Prediction Network with Multi-Level Graph Contrastive Learning

**Authors:** Yang An, Bo Jin, Xiaopeng Wei  
**arXiv:** [2204.11736](https://arxiv.org/abs/2204.11736)  
**Date:** 2022-04-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2204.11736v2/

## Key points

- Predicting medications in intelligent healthcare systems is crucial for assisting doctors in making informed decisions based on patients' electronic medical records (EMRs).
- Existing studies have focused on utilizing relations between homogeneous codes from medical ontology graphs to enhance predictions using supervised methods.
- Few studies have explored valuable relations between heterogeneous or homogeneous codes from historical EMRs, limiting prediction performance and application scenarios.
- KnowAugNet is introduced as a multi-sourced medical knowledge augmented medication prediction network that captures diverse relations between medical codes through a multi-level graph contrastive learning framework.
- KnowAugNet utilizes graph attention networks and weighted graph convolutional networks to capture implicit and correlative relations between medical codes from both medical ontology graphs and constructed medical prior relation graphs.

## Summary

, , , , 

In the field of intelligent healthcare systems, predicting medications is crucial for assisting doctors in making informed decisions based on patients' electronic medical records (EMRs). However, this task is challenging due to the complex relationships between medical codes. Existing studies have focused on utilizing relations between homogeneous codes from medical ontology graphs to enhance predictions using supervised methods. However, few have explored the valuable relations between heterogeneous or homogeneous codes from historical EMRs, limiting prediction performance and application scenarios. To address these challenges, this paper introduces KnowAugNet, a multi-sourced medical knowledge augmented medication prediction network that captures diverse relations between medical codes through a multi-level graph contrastive learning framework. KnowAugNet utilizes graph attention networks and weighted graph convolutional networks to capture implicit and correlative relations between medical codes from both medical ontology graphs and constructed medical prior relation graphs. By leveraging these relationships, KnowAugNet can generate augmented medical code embedding vectors that are input into a sequential learning network to predict medications for patients while considering temporal relations. The paper defines clinical patient records as longitudinal sequential history medical records containing multivariate observations for each patient. It also introduces the concept of multi-sourced medical knowledge graphs, including common-sense medical ontology graphs and empirical medical prior relation graphs. Experimental results evaluating KnowAugNet's performance demonstrate its effectiveness in medication prediction. The paper discusses experiment details such as dataset selection and preprocessing, evaluation metrics, benchmark models, and experimental settings. It also analyzes prediction results, conducts an ablation study on model components, explores the impact of graph sparsity factors in medical prior relation graphs, and assesses the effects of graph encoders in graph contrastive learning frameworks. Overall,<OrganizationName> shows promise in improving medication prediction accuracy by leveraging diverse relationships between medical codes from various sources. The paper concludes with discussions on limitations and future research directions in this area of study.