The Graph-based Mutual Attentive Network for Automatic Diagnosis

Quan Yuan, Jun Chen, Chao Lu, Haifeng HuangIJCAI 2020 · DOI

Introduction

Automatic diagnosis systems aim to predict diseases from Electronic Medical Records (EMRs). Despite advances in deep learning, three primary challenges persist: the complexity of medical relationships, feature sparsity due to varied clinical writing styles, and the loss of contextual nuance when focusing solely on extracted medical entities. Traditional sequence models often fail to capture the structural "web" of symptoms and diseases or the critical temporal context (e.g., "ten years" vs. "one hour") hidden in non-medical keywords.

Methodology

The Graph-based Mutual Attentive Network (GMAN) addresses these challenges through a three-stage process: graph construction, GCN encoding, and mutual attention.

1. Medical Graph Construction

To incorporate domain expertise, GMAN utilizes two structured graphs: a Disease Hierarchy based on ICD-10 codes and a Causal Graph mined from historical EMR data. These graphs map "is-a" relationships and link diseases to their common symptomatic findings, providing the model with a form of medical "common sense."

The medical graph structure incorporating disease hierarchy and causal relationships between diseases and symptoms.

2. GCN Encoding

To mitigate feature sparsity, a Graph Convolutional Network (GCN) is employed. Instead of treating medical terms as isolated features, the GCN updates symptom embeddings based on their neighbors in the medical graph. This ensures that synonymous or related terms share similar mathematical representations, enhancing the model's robustness to linguistic variation.

3. Mutual Attentive Network

The core reasoning engine is a two-step consecutive attention mechanism that bridges the gap between structured findings and unstructured text. Unlike parallel models, GMAN's consecutive approach allows medical entities to deeply inform the interpretation of raw text.

The architecture of GMAN, illustrating the flow from EMR data through GCN encoding and the Mutual Attentive Network.
  • Text-Guided Attention: The model uses the global context of the clinical notes to weight the importance of extracted medical findings.
  • Finding-Guided Attention: It then uses these weighted findings to re-examine the original text, highlighting critical contextual words (like symptom duration) that might have been missed during initial entity extraction.

Evaluation and Results

GMAN was evaluated on three large-scale datasets: proprietary Chinese Neurology and Cardiology datasets (36,000+ samples) and the public English MIMIC-III-50 benchmark.

The model consistently outperformed standard baselines (CNN, BiGRU) and specialized medical models (ACNN, CAML):

  • Neurology: Achieved a 73.59% Recall@1, a ~9% absolute improvement over the best baseline.
  • Cardiology: Achieved a 57.31% Recall@1, outperforming others by roughly 4-5%.
  • MIMIC-III-50: Established a new high bar with a 63.46% Precision@5.

Beyond accuracy, the model offers interpretability by using attention weights to highlight which symptoms and specific words led to a diagnosis, a crucial feature for clinical trust.

Conclusion

By integrating structured medical knowledge with deep learning on unstructured text, GMAN provides a more accurate and interpretable approach to automatic diagnosis. Its successful deployment in hundreds of primary care facilities in China demonstrates its real-world utility. Future work will explore the integration of laboratory values and imaging data to further refine diagnostic accuracy.

Summary updated

BibTeX
@inproceedings{yuan2020graph,
  author = {Quan Yuan and Jun Chen and Chao Lu and Haifeng Huang},
  title = {{The Graph-based Mutual Attentive Network for Automatic Diagnosis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/469},
}
Figure