Automatic Emergency Diagnosis with Knowledge-Based Tree Decoding

Ke Wang, Xuyan Chen, Ning Chen, Ting ChenIJCAI 2020 · DOI

TL;DR

This paper introduces K-BTD, a novel knowledge-based tree decoding model for automatic emergency diagnosis. By leveraging the hierarchical structure of ICD codes and external medical knowledge, K-BTD achieves state-of-the-art performance and offers a more interpretable, step-by-step diagnostic reasoning process compared to existing methods.

What Problem Does This Paper Address?

Accurate and timely diagnosis is crucial in emergency departments, but manual diagnosis is time-consuming and prone to errors. Existing automatic diagnosis systems often treat diseases as independent entities and fail to utilize the inherent hierarchical structure of medical coding systems like ICD. Furthermore, they lack transparency in their decision-making, making it difficult for clinicians to trust and verify their outputs. This paper aims to bridge this gap by developing a model that incorporates the ICD tree structure and external medical knowledge to provide a more accurate, interpretable, and human-like diagnostic process.

Methods

The core of the proposed method is the Knowledge-Based Tree Decoding (K-BTD) model. The inference process is a top-down traversal of the ICD tree, mimicking a human doctor's diagnostic approach. Key components include:

  • Clinical Notes Encoder: Processes patient clinical notes using an embedding layer and an RNN (LSTM, GRU, or BERT) to extract contextual features.
  • Knowledge Encoder: Encodes external medical knowledge (from Wikipedia/Baidupedia) associated with each ICD node using CNNs and a self-attention mechanism.
  • Judge Net: At each node in the ICD tree, this component decides whether to expand a child node for further decoding. It uses two scores:
  • Semantic-level score: Captures interactions between the clinical notes' aggregated information and the child node's external knowledge.
  • Word-level score: Captures word-level interactions between clinical notes and the child node's knowledge. An overall score determines if a child node is expanded.

  • Fusion Net: If a child node is chosen for expansion, this component fuses the parent node's flowing vector with the child node's external knowledge to generate the child node's flowing vector.

This stepwise decoding process allows the model to provide support for each decision, enhancing interpretability.

Datasets

The study uses a dataset collected from the emergency department of Beijing Tsinghua Changgung Hospital, spanning from 2015 to 2017. This dataset contains unstructured clinical notes and corresponding ICD-9 codes. For external medical knowledge, Wikipedia and Baidupedia were utilized. The experiments focused on the top-100 and top-150 most frequent ICD-9 codes to manage the large taxonomy.

Key Insights and Ideas

  • Leveraging Hierarchical Structure: The paper strongly advocates for the importance of the ICD tree structure. By treating diagnosis as a tree-decoding problem, the model can enforce medical relationships between diseases, preventing illogical diagnoses.
  • External Knowledge Integration: Incorporating external medical knowledge from sources like Wikipedia and Baidupedia significantly enhances diagnostic accuracy and provides richer context.
  • Stepwise Interpretability: The top-down decoding process allows for "stepwise support," where the model highlights the evidence from clinical notes and external knowledge at each decision point. This makes the diagnostic reasoning transparent and more useful for clinicians.
  • Knowledge Quality Matters: Ablation studies reveal that the model's performance is sensitive to the quality of external knowledge, with noise significantly degrading accuracy.

Key Results

  • State-of-the-Art Performance: K-BTD outperforms all baselines in both micro-F1 and macro-F1 scores across experiments on top-100 and top-150 ICD codes. For example, on top-100 codes, K-BTD (BERT) achieved 0.7085±0.0128 micro-F1 and 0.5973±0.0097 macro-F1.
  • Reduced Semantic Distance: The model significantly reduces semantic distance, indicating its predictions are closer to the ground truth from a medical perspective.
  • Improved Interpretability: Qualitative evaluations demonstrate that K-BTD's stepwise support provides more intuitive and actionable insights compared to single-step attention models.

Limitations and Future Work

  • Focus on Frequent Diseases: The current study primarily focuses on high-frequency ICD codes. Future work aims to increase the diagnosis accuracy for infrequent diseases.
  • External Knowledge Dependency: The model's performance is heavily reliant on the quality of external medical knowledge. Future research could explore more robust methods for knowledge extraction and integration.
  • Model Architecture: While K-BTD is presented as a framework, further exploration of different text encoders (e.g., more advanced transformer models) could lead to even higher performance.

Summary updated

BibTeX
@inproceedings{wang2020automatic,
  author = {Ke Wang and Xuyan Chen and Ning Chen and Ting Chen},
  title = {{Automatic Emergency Diagnosis with Knowledge-Based Tree Decoding}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/471},
}

Loading related content...

Figure